Evaluating and improving indoor positioning methods(3)WSN optimization deployment and generation of radio{maps which can be implemented in IoT and WSNs localization especially in indoor

Evaluating and improving indoor positioning methods

Citation for published version (APA):Kanaris, L. (2019). Evaluating and improving indoor positioning methods. Technische Universiteit Eindhoven.

Document status and date:Published: 10/12/2019

Document Version:Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers)

Please check the document version of this publication:

• A submitted manuscript is the version of the article upon submission and before peer-review. There can beimportant differences between the submitted version and the official published version of record. Peopleinterested in the research are advised to contact the author for the final version of the publication, or visit theDOI to the publisher's website.• The final author version and the galley proof are versions of the publication after peer review.• The final published version features the final layout of the paper including the volume, issue and pagenumbers.Link to publication

General rightsCopyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright ownersand it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.

• Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portal.

If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, pleasefollow below link for the End User Agreement:www.tue.nl/taverne

Take down policyIf you believe that this document breaches copyright please contact us at:[email protected] details and we will investigate your claim.

Download date: 17. Aug. 2020

https://research.tue.nl/en/publications/evaluating-and-improving-indoor-positioning-methods(856c392b-44d8-42f6-8675-dc7e2d25697c).html

Evaluating and Improving IndoorPositioning Methods

PROEFSCHRIFT

ter verkrijging van de graad van doctor aan de Technische UniversiteitEindhoven, op gezag van de rector magnificus prof.dr.ir. F. Baaijens voor eencommissie aangewezen door het College voor Promoties, in het openbaar te

verdedigen op dinsdag 10 december 2019 om 11:00 uur

door

Loizos Kanaris

geboren in Nicosia, Cyprus

Dit proefschrift is goedgekeurd door de promotoren en de samenstelling van depromotiecommissie is als volgt:

Voorzitter: prof.dr.ir. A.M.J. Koonen1e promotor: prof.dr.ir. A. Liotta2e promotor: prof.dr.ir. S. Stavroucopromotor(en): dr.ir. G. Exarchakosleden: dr.ir. T. Dagiuklas (London South Bank University)

prof.dr.ir. S. Kotsopoulos (University of Patras)prof.dr.ir. PG.M. Baltus

Het onderzoek of ontwerp dat in dit proefschrift wordt beschreven is uitgevoerd inovereenstemming met de TU/e Gedragscode Wetenschapsbeoefening.

Copyright © 2019 Loizos Kanaris, Eindhoven, the Netherlands.All rights reserved. A catalogue record is available from the Eindhoven Universityof Technology Library.

ISBN: 978-90-386-4924-5NUR: 959

Acknowledgements

I would like to express my gratitude to the management of Sigint SolutionsLtd for sponsoring this PhD. The unlimited access to their ISO 17025 laboratorypremises and equipment, as well to the free utilization of their software tools, likeTruNET wireless radio planning simulator, was of critical importance during theexperimentation and testing of novel algorithms and methodologies introduced inthe published scientific work.

I also thank my supervisors, prof.dr. Antonio Liotta and prof.dr. Stavros Stavrou,for their wise advises, out of the box thinking and restless support.

Finally, it would be an omission not to acknowledge the emotional support andpatience of my family throughout the whole period of my PhD.

Loizos KanarisEindhoven, The Netherlands, 2019

i

Summary

Evaluating and Improving Indoor Positioning Methods

User positioning in indoor areas, where Global Navigation Satellite Systems (GNSS)are not available, has been a challenge to the research community for the lasttwo decades. The increased demand raised by numerous industries for more ac-curate, reliable and real time user or asset location information, has been driv-ing researchers to keep investigating various, off-the-shelf positioning approaches.Within this concept, researchers tend to utilize and combine existing and newwireless communication technologies such as cellular, Wi-Fi, Bluetooth, Zigbee,VLC etc., or fuse any useful available information in order to improve the speedor accuracy of the positioning process. Indoor positioning methods involve thecombination of different parameters such as Time of Flight (TOF), Received Sig-nal Strength (RSS), RSS Difference (RSSD), Direction of Arrival (DOA), Time OfArrival (TOA), Time Difference Of Arrival (TDOA), etc. Among all, fingerprint-based positioning is one of the most popular indoor localization techniques im-plemented by Real Time Localization Systems (RTLS), because of its deploymentsimplicity and low cost. Positioning systems that are based on fingerprint maps, orother relevant methods, can find use in malls, for providing indoor navigation andgeo-marketing services; in hospitals, for monitoring patients, doctors and criticalequipment movement; in logistics, for tracking assets and optimizing empty spacesin ports or in land storage facilities; and in homes, for providing Ambient AssistedLiving (AAL) services.

One aspect of the challenges faced in all types of indoor positioning systems isto objectively evaluate the performance of the approaches and methodologies im-plemented. At the same time, being in the era of heterogeneous technologies co-existence and computational booming, various possibilities of information fusioncan be investigated. The aforementioned observations have been the motivationto set a twofold research scope in this thesis:

(1) To investigate the existing evaluation methodologies related to IndoorPositioning Systems and platforms, identify possible weak areas and pro-pose new techniques and algorithms that will contribute towards a moreobjective and unbiased performance evaluation of existing, or newly pro-posed indoor positioning methods and algorithms.

(2) To propose new indoor positioning algorithms with the aim to improvethe indoor localization accuracy achieved, by combining data from dif-ferent technologies or sources.

iii

SUMMARY

Towards the evaluation of the indoor positioning methodologies, which is themain direction of the research work, we introduced 3 novel algorithms:

(1) Sample Size Determination Algorithm (SSDA) for the evaluation of fingerprint-based indoor localization systems, as described in Chapter 3 ([1]).

(2) Tolerance Based - Normal Probability Distribution (TBNPD) correlationalgorithm for the quality evaluation of any fingerprint-based dataset, aspresented in Chapter 4 ([2]).

(3) A Binomial Distribution Approach for the evaluation of already deployedindoor positioning systems that is analyzed in Chapter 5 ([3]).

On the second aspect of this thesis, which involved the combination of dataand technologies for improving indoor localization accuracy, several new methodswere tested and proposed as per Chapter 6:

(1) Combining BLE and Wi-Fi technologies ([4]).(2) Combining smart lighting and radio technologies ([5]).(3) Utilizing Directional Antennas for improving a radiomaps’ quality char-

acteristics ([6]).

In the same scope, further contributions were made in:

(1) Map-aided localization methods, where positioning accuracy improve-ment was achieved by imposing map-constraints into the positioning al-gorithms in the form of a–priori knowledge ([7] and [8]).

(2) Single Access Point positioning, by exploring polarization in MIMO de-vices (publication submitted).

(3) WSN optimization deployment and generation of radio–maps which canbe implemented in IoT and WSNs localization especially in indoor envi-ronments ([9]).

(4) Self-Organization Networks, where the generated radiomaps and the lo-cation of the mobile users are utilized towards reduced cost and energyper bit for future emerging radio technologies, i.e. 5G networks ([10]).

iv

Contents

Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i

Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii

List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi

Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1. Overview of Positioning Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1.1. Outdoor Positioning Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.1.2. Indoor Positioning Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.2. Research Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81.3. Structure of Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2. Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112.2. Positioning Techniques. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112.3. Fingerprint-based Positioning Methodologies . . . . . . . . . . . . . . . . . . . . . . . 212.4. Hybrid indoor RTLS and sensor information fusion. . . . . . . . . . . . . . . . . 262.5. RTLS Performance Evaluation Techniques . . . . . . . . . . . . . . . . . . . . . . . . . 282.6. The Challenge of Selecting the Right Testing Sample . . . . . . . . . . . . . . . 29

3. Sample Size Determination for the Evaluation of Indoor LocalizationSystems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333.2. Proposed Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333.3. Test Environment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373.4. Performance Evaluation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383.5. Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

4. Quality Evaluation of Fingerprint–based RTLS . . . . . . . . . . . . . . . . . . . . . . . . . . 474.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474.2. Proposed Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474.3. Test Environment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554.4. Performance Evaluation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 564.5. Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

5. Evaluation of already deployed Positioning Systems . . . . . . . . . . . . . . . . . . . . . 63

v

CONTENTS

5.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 635.2. Proposed Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 635.3. Test Environment and Measurement Setup. . . . . . . . . . . . . . . . . . . . . . . . . 655.4. Performance Evaluation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 675.5. Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

6. Fusing Information for Improving Indoor Positioning Accuracy . . . . . . . . . . 716.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 716.2. Combining IEEE 802.15 (BLE) and IEEE 802.11 (Wi-Fi)

Technologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 716.3. Combining Smart Lighting and Radio in Indoor Positioning . . . . . . . . 806.4. Directional Antennas for Improving Indoor Positioning . . . . . . . . . . . . . 866.5. Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

7. Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 937.1. Conclusion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 937.2. Future Work. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

List of Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

Curriculum Vitae . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

vi

List of Figures

Fig. 1.1 GNSS Concept 3Fig. 1.2 Indoor Positioning Systems Categories 6

Fig. 2.1 Typical TOA Positioning Concept 12Fig. 2.2 Typical TDOA Positioning Concept 14Fig. 2.3 Typical DOA Positioning Concept 15Fig. 2.4 Fingerprinting: Offline Phase 18Fig. 2.5 Fingerprinting: Online Phase 19Fig. 2.6 Dead Reckoning Concept [source: PNI Sensor Corp. (Santa Rosa,

Calif.)] 21Fig. 2.7 Fingerprint-Based Localization Concept 22

Fig. 3.1 t-distribution table sample 35Fig. 3.2 Test Environment with APs and Radiomap as Generated in

TruNET Simulator 37Fig. 3.3 Positioning Error Vectors-Sample of Convenience con5 with

ncon = 10 - WKNN 39Fig. 3.4 Positioning Error Vectors-Sample of Convenience con5 with

ncon = 10 - MMSE 39Fig. 3.5 Positioning Error Distribution-Sample of Convenience con5 with

nsps = 10 40Fig. 3.6 Positioning Error Vectors-Random Sample sps3 with nsps = 10 -

WKNN 40Fig. 3.7 Positioning Error Vectors-Random Sample sps3 with nsps = 10 -

MMSE 41Fig. 3.8 Positioning Error Distribution-Random Sample sps3 with

nsps = 10 41Fig. 3.9 Radiomap generated through measurements: Effect of sample size

on positioning error for WKNN 43Fig. 3.10 Radiomap generated through measurements: Effect of sample size

on positioning error for WKNN 43Fig. 3.11 Typical Simple Random Sample Selection of 48 locations for

WKNN Algorithm - Small Preliminary Sample No3 44Fig. 3.12 radiomap generated by a high-resolution simulation: Effect of

sample size on positioning error using WKNN algorithm 44

Fig. 4.1 RET: Main Window Toolbar - Part 1 51Fig. 4.2 RET: Main Window Toolbar - Part 2 51

vii

CONTENTS

Fig. 4.3 RET: Selected Location vs All 53Fig. 4.4 Graphical Presentation of radiomap Signatures 53Fig. 4.5 3x3 Evaluation Grid for the radiomap Signature 54Fig. 4.6 Comparison of radiomap quality: signature shift to low left: LOW

Correlation - LOW Error 54Fig. 4.7 Test Environment: Measurement points and APs. 55Fig. 4.8 Radiomap Signature for 3 APs 57Fig. 4.9 Radiomap Signature for 6 APs 57Fig. 4.10 Effect of APs number: MMSE Algorithm 58Fig. 4.11 Effect of APs number: WKNN Algorithm 58Fig. 4.12 Equilateral and Obtuse Triangular Deployment 59Fig. 4.13 radiomap Signature for 3 APs - Areat1 60Fig. 4.14 radiomap Signature for 3 APs - Areat2 60Fig. 4.15 Empirical and Simulated radiomap Signatures: Effect of RSStol

±2dB 61Fig. 4.16 Empirical and Simulated radiomap Signatures: Effect of RSStol

±6dB 61Fig. 4.17 Effect of RSS Tolerance at the Output Level: WKNN and MMSE

- 6 APs Active 62

Fig. 5.1 Sample of z-score table 65Fig. 5.2 Measurement points and APs in the Laboratory Premises 66Fig. 5.3 RSS Correlation Plot between Galaxy Tab and Nexus before LT 69Fig. 5.4 RSS Correlation Plot between Galaxy Tab and Nexus after LT 69

Fig. 6.1 Concept of i-KNN: BLE utilization for Radiomap SubsetGeneration 73

Fig. 6.2 φ−map Localization Platform 76Fig. 6.3 Positioning Error Comparison: Wi-Fi only vs single BLE and

Wi-Fi 79Fig. 6.4 Fingerprint Dataset Size Utilization: Wi-Fi only vs single BLE

and Wi-Fi 79Fig. 6.5 Positioning Error Comparison: Wi-Fi only vs nearest BLE and

Wi-Fi 79Fig. 6.6 Fingerprint dataset Size Utilization: Wi-Fi only vs nearest BLE

and Wi-Fi 80Fig. 6.7 Angles φ and θ to compute channel impulse response and SL

Radius 81Fig. 6.8 Combined VLP and Wi-Fi Fingerprint Based Indoor Positioning 83Fig. 6.9 Positioning Error Comparison: WiFi only vs VLP-WiFi 86Fig. 6.10 Fingerprint Dataset Size Utilization: Wi-Fi only vs VLP-WiFi 86Fig. 6.11 Test Environment: Measurement points and APs. 88Fig. 6.12 Directional Antenna: Vertical and Horizontal Cut Patterns. 88Fig. 6.13 Positioning Error per Azimuth 89Fig. 6.14 Error Improvement: Omni Vs. Omni & Directional Antennas 90Fig. 6.15 Correlation Level Improvement: Omni Vs. Omni & Directional

Antennas 90

viii

CONTENTS

Fig. 6.16 Radio coverage of tested Radio–maps: Maximum RSS values perlocation. 91

ix

List of Tables

Tab. 1.1 GNSS Positioning System Comparison 4

Tab. 1.2 Cellular Networks Localization methods and accuracy 5

Tab. 1.3 Indoor Positioning:Technologies comparison 7

Tab. 1.4 Indoor Positioning: Typical Active Localization Systems 8

Tab. 3.1 Material Constitutive Parameters of the Test Environment 38

Tab. 3.2 Mean Error and Standard Deviation of Preliminary Samples-radiomap generated through Measurements 42

Tab. 3.3 Suggested Sample Size for different desired EB values using Sampleof Convenience– radiomap generated through Measurements 42

Tab. 3.4 Suggested Sample Size for different desired EB values using SimpleRandom sps –radiomap generated through Measurements 42

Tab. 4.1 Measurement Sample and Measurement Profile Parameters 49

Tab. 4.2 Transmitter and Device Parameters 50

Tab. 4.3 Material Constitutive Parameters of the Test Environment 56

Tab. 4.4 Radiomap Quality Evaluation using TBNPD Algorithm: Numberof APs: 3 APs vs. 6 APs at RSStol = ±2dB 58

Tab. 4.5 Equilateral vs. Obtuse Triangular Deployment 60

Tab. 4.6 radiomap Quality Evaluation using TBNPD Algorithm: RSSTolerance Effects - 6 APs Active 61

Tab. 5.1 Continuity Correction Table 65

Tab. 5.2 Calculated Accuracy per Device 66

Tab. 5.3 Outcome of Binomial Approach for the KNN Algorithm 68

Tab. 5.4 Outcome of Binomial Approach for the MMSE Algorithm 68

Tab. 5.5 Outcome of Binomial Approach for Calibrated Radiomaps 70

Tab. 6.1 Positioning Error of Wi-Fi RSS Fingerprint-based PositioningSystem 76

Tab. 6.2 Positioning Error of Typical Indoor Wi-Fi Positioning Systems 77

Tab. 6.3 Positioning Error of Combined BLE (single BLE) and Wi-Fi RSSFingerprint-based Positioning System 78

Tab. 6.4 Positioning Error of Combined BLE (all deployed BLEs) andWi-Fi RSS Fingerprint-based Positioning System 78

Tab. 6.5 Positioning Error of Wi-Fi RSS Fingerprint-based PositioningSystem 85

xi

CONTENTS

Tab. 6.6 Positioning Error of Combined VLP and Wi-Fi RSS Fingerprint-based Positioning System 85

Tab. 6.7 Positioning Error per Azimuth - WKNN 89Tab. 6.8 Positioning Error per Azimuth - MMSE 89

xii

Abbreviations

Abbreviation DescriptionAAL Ambient Assisted LivingAR-WFL Aging-Resistant Wi-Fi Fingerprinting LocalizationAOA Angle of ArrivalAP Access PointBDS BeiDou Navigation Satellite SystemBLE Bluetooth Low EnergyBSN Body Sensor NetworksCEP Circular Error ProbabilityCDF Cumulative Distribution FunctionD2D Device to DeviceDAS Delay and SumDCM Database Correlation MethodDOA Direction of ArrivalDOAJ Directory of open access journalsE-CID Enhanced Cell IDE-OTD Enhanced Observed Time DifferenceESPRIT Estimation of Signal Parameters via Rotational In-

variance techniqueGFSK Gaussian Frequency Shift KeyGLONASS Global Navigation Satellite SystemGPS Global Positioning SystemGO Geometrical OpticsGTD Geometrical Theory of DiffractionHLF Hyperbolic Location Fingerprintingi-KNN intelligent KNNIMU Inertial Measuring UnitIoT Internet of ThingsIW-KNN Iterative Weighted KNNKNN K Nearest NeighbourLBS Location Based ServicesLED Light-Emitting DiodeLoB Line of bearingLT Linear TransformationLTE Long Term EvolutionMAP Maximum A PosterioriMDPI Multidisciplinary Digital Publishing Institute

xiii

ABBREVIATIONS

MMSE Minimum Mean Square ErrorMS Mobile StationMU Mobile UserMUSIC Multiple Signal ClassificationNMF Neighbor Mean FilterNN Nearest NeighbourOTDoA Observed Time Difference of ArrivalOWC Optical Wireless CommunicationRSS Received Signal StrengthRSSD RSS DifferenceRT Ray TracingRTLS Real Time Localization SystemSDR Software Defined RadioSSDA Sample Size Determination AlgorithmSON Self Organizing NetworksTDOA Time Difference of ArrivalTOA Time of ArrivalTOF Time of FlightUE User EquipmentUTD Uniform Theory of DiffractionUToA Uplink TOAUTDoA Uplink Time Difference of ArrivalUWB Ultra Wide BandVLC Visible Light CommunicationVLP Visible Light PositioningVANET Vehicular Ad-hoc NetworksWFL Wi-Fi Fingerprint-based localizationWKNN Weighted K Nearest Neighbour

xiv

CHAPTER 1

Introduction

1.1. Overview of Positioning Systems

Localization or positioning of users and assets, refers to the process of lo-cating and tracking them via wireless devices, carried by users, or attached toassets. Research on positioning techniques, has been conducted for decades inboth outdoor and indoor environments, with the military and security organiza-tions leading the way. More broad interest became obvious, with the introductionof cellular services and the requirement from operators to provide user locationinformation to public safety services in emergency situations [11]. Nowadays withthe rapid advances in Wireless Communication Systems, a wide range of position-ing applications has evolved in different industry sectors, involving both personand asset tracking. Such applications expand much further than the typical mili-tary tracking of friend and foe forces. They include the medical sector for locatingdoctors and patients, logistics industry to track and support assets and equip-ment, as well as in marketing for advertising products to target groups, basedon their geographical location [12]. A more advanced series of applications canfurther provide location-sensitive billing, cellular phone fraud detection, as wellas navigation from and to almost everywhere, through the utilization of hetero-geneous wireless technologies, fusion of sensor and IoT data ([13], [14], [15], [16]).

In order to investigate how to accurately and reliably provide positioning ser-vices everywhere, researchers examined the nature of the wireless channel withan aim to estimate meaningful parameters and techniques that could be used forlocalization purposes. The most frequently employed techniques, for providingindoor and outdoor localization services include:

(1) Range based lateration, which requires measurements of the distanceof the user from at least three known locations in a 2-D space. Distancemeasurements can be achieved either through the utilization of ReceivedSignal Strength (RSS) or the Time of Arrival (TOA) or the Time Differ-ence of Arrival (TDOA).

(2) Direction based angulation that involves the calculation of the Di-rection of Arrival (DOA) - or the angle of arrival (AOA) - of at least twosignals from known locations.

(3) Fingerprinting, where the location is identified by comparing the signalcharacteristics of pre-collected data with the characteristics of the usersignals in the area of interest.

(4) Dead reckoning which calculates the position of the user, with referenceto fixed known positions, using distance, speed and/or drift angle data.

1

CHAPTER 1. INTRODUCTION

All the above can of course be combined with, or assisted by, non-radio param-eters - such as maps, or inertial sensors, in order to provide enhanced positioningservices. Details on wireless indoor positioning techniques and systems can befound in [17], [18] and [19].

Having stated the above, one of the core ongoing challenges in the area, is toseek for approaches that will improve accuracy, while at the same time be afford-able to the wide market. This led the research community towards the utilizationof existing of-the shell equipment and devices, combined with fusion of informationretrieved by available sensors. In all cases, localization accuracy depends on theunderlying technologies, the methods used and the complexity and dynamism ofthe varying radio channel.

Analyzing the localization problem from a high level perspective, this can be sep-arated based on the environment. Depending if one is outdoor or indoor, differenttechnologies can be used for localization principles. In the latter, the complexityincreases due to multipath effects which are highly dependable on the constitutiveparameters of the building materials interacting with the radio waves and the in-door geometry of the premises.

For this reason, a brief overview on outdoor and indoor localization systems isprovided in this chapter, in subsections 1.1.1 and 1.1.2 respectively. The researchscope is defined in section 1.2 and the structure of the thesis is explained in section1.3.

1.1.1. Outdoor Positioning Systems

1.1.1.1. Global Navigation Satellite Systems. When it comes to localizationin outdoor environment, a huge advantage exists, when compared to the indoorenvironment: The possibility to have access and utilize satellite positioning tech-nologies. The US Global Positioning System (GPS), the Russian Global NavigationSatellite System (GLONASS), the European GALILEO and the Chinese BeiDouNavigation Satellite System (BDS) are currently the dominant Global NavigationSatellite Systems (GNSS) providing positioning services in outdoor environments[20]. All systems utilize satellite constellations, that orbit at predetermined heightsand on several orbital planes, in such a way that, at the same time, several satel-lites provide positioning and timing information within the coverage area.

Two more segments, in addition to a satellite constellation, complete a GNSSsystem: the ground control network, and the user equipment.The localizationtechnique utilized in this category of systems is Time–based technique, and forthis reason high accuracy clocks are used for synchronization of the satellites andthe ground control stations. In order to minimize the error of the receiver clocks,which are not as accurate as the satellite clocks, information is required from 4satellites in 3D-space. Conceptually, there are four steps during the localizationprocedure in GNSS as shown in figure 1.1 and summarized below:

2

1.1. OVERVIEW OF POSITIONING SYSTEMS

Figure 1.1 – GNSS Concept

(1) Satellite Constellation: Satellites orbit on their orbital planes equippedwith high accuracy clocks and knowing their orbit ephemerides (the pa-rameters that define their orbit). Ground-based control stations adjustthe satellites ephemerides and time, when necessary.

(2) Propagation: Each satellite broadcasts on a regular basis its status,ephemerides and time. Radio signals pass through layers of the atmo-sphere to the user equipment.

(3) Reception: Users receive the signals from multiple GNSS satellites.Each satellite provides the necessary information for calculation of thepropagation time.

(4) Computation: User equipment uses the received information for timeand position computation.

(5) Application: The User application installed on the equipment, utilizesthe computed time and position information to provide the service re-quested, ie navigation, mapping, etc.

The range of envisaged applications of GNSS systems is enormous, spanningfrom the military, public and private sectors across numerous market segments.Typical examples, apart from the conventional LBS and navigation services, in-clude the provision of timing, positioning and velocity in IoT and context awareapplications, emergency, security and humanitarian services, science, environment,weather, agriculture and much more [21].

The characteristics and performance data of the most important GNSS sys-tems [22], is summarized in table 1.1.

1.1.1.2. Cellular Networks for Localization Services. Cellular systems are cur-rently in every country providing almost total coverage over the land, to anycellular device. The introduction of the requirement for location information into

3


Table 1.1 – GNSS Positioning System Comparison

Parameter GPS GLONASS GALILEO Beidou

1st Launch 22-Feb-78 12-Oct-82 28-Dec-05 13-Apr-07Services mil-civil mil-civil com-open auth-openIn orbit Sat 31 26 28 23Max Sat 33 24 30 35Orbital Planes 6 3 3 3Inclination 55° 64.8° 56° 55°Orbit Alt[km] 20.183 19.100 23.222 21.500Orbital Period 11h55m 11h15m 14h05m 12h50mCoverage [km2] 193.705.518 191.117.254 199.969.523 196.773.430Time System GPST UTC(SU) GST UTC (China)Coding CDMA FDMA CDMA CDMAFrequencies L1,L2,L5 G1,G2,G3 E1,E5a,E5b,E6 B1-2,B1,B2,B3Accuracy [m] 5 2.8-7.38 1 10

standardization, was triggered after the decision of Federal Communications Com-mission (FCC) of the United States (U.S.) which defined enhanced 911 (E911)location requirements in the mid 1990s. In the same line, in 2000, the Euro-pean Commission (EC) formed, in the framework of harmonizing emergency ser-vices, a coordination group on access to location information by emergency services(CGALIES). Both governmental institutions demanded to locate mobile terminalsin case of an emergency. Setting as a driving factor the need for user localization incase of emergencies, providers have gradually introduced Location Based Servicesin all cellular technologies, from 1G to 5G. A comprehensive analysis of cellularmobile localization methods is provided in [23].

In general, cellular networks utilize the localization techniques based on TOA,triangulation, and proximity. Based on the concept of these techniques the follow-ing methods can be implemented: UToA (Uplink TOA), UTDoA (Uplink TimeDifference of Arrival) OTDoA (Observed Time Difference of Arrival) E-CID (En-hanced Cell ID), TA/CID, E-OTD (Enhanced Observed Time Difference) etc.Additionally, outdoor fingerprinting, known also as Scene Analysis can be im-plemented [24].

The achieved localization accuracy per method and technology in cellular net-works is summarized in table 1.2 [23], [25], [24] and [26].

The location of a mobile device can be used to provide additional functionali-ties to the user, resulting in the so-called location-based services (LBS). In additionto the LBS, the location information is utilized for improving the communicationcapacity, by supporting decisions for better network management and for a moreoptimized re-configurable radio spectrum. User location is also important for ap-plications in intelligent transportation systems (ITS), vehicular ad-hoc networks(VANETs), resource management in device-to-device (D2D) communications, etc.This type of localization applications can be also added to the self-organizingnetworks (SON). SON is a mechanism to ease and improve the deployment and

4


Table 1.2 – Cellular Networks Localization methods and accuracy

Method Technology Type 67%-Accuracy[m]

CID+TA/CID+RTT 2G,3G,4G Proximity > 100E-CID 3G,4G Proximity ∼ 50RFPM 2G,3G,4G Fingerprinting > 50U-ToA 2G Trilateration > 100UTDoA 3G,4G Trilateration > 50E-OTD 2G Trilateration > 100AFLT 3G Trilateration > 50OTDoA 3G, 4G Trilateration ≤ 50TBS 4G Trilateration ≤ 50

operation of a network, exploited commercially by the network operator or theapplication developer, with an aim to obtain a revenue [10].

1.1.2. Indoor Positioning Systems

For environments where GNSS signals are not available (indoor and Manhattan-like outdoor environment), there is a need for utilizing other wireless technologies inorder to provide location based services. Numerous positioning methods have beeninvestigated by the research community, and many of them have been adopted bythe industry. Indoor localization systems may combine homogeneous or heteroge-neous Wi-Fi, WiMAX, LTE, Zigbee, UWB technologies. They typically implementwell known techniques utilizing RSS, TOA, TDOA and DOA parameters. Severalextensive surveys were performed in the positioning area for categorizing and com-paring the performance of the proposed approaches, experimental platforms andsystems. In [27] a categorization is presented based on whether a technique isbuilding dependant or not and whether there is a need of infrastructure utiliza-tion. An updated version of the same approach is presented in figure 1.2. Theadvantages and disadvantages of the technologies involved are summarized in table1.3.

In another approach, authors in [34], enumerated various active and pas-sive localization techniques and they assess them with respect to accuracy andprecision, complexity, scalability and cost. In [35] the researchers collected andcompared recent emerging fingerprinting solutions that try to handle the intenseeffort required during the offline phase. An updated list of typical, well knownactive localization systems is presented in table 1.4.

During the last two decades, the research community has concentrated on theutilization of IEEE 802.11 technology for indoor location estimation, due to itspopularity and dominance of its devices in the global market. Nowadays, RSSfingerprint-based positioning techniques are often applied in various indoor local-ization scenarios, because of their implementation simplicity and their ability to

5


Figure 1.2 – Indoor Positioning Systems Categories

provide acceptable accuracy. An overview of the different fingerprint-based meth-ods is provided in [36].

The most recent advances on Wi-Fi fingerprint localization are presented in [37].Authors studied the advances by separating them into two major areas: (i) Ad-vanced localization techniques and (ii) Efficient system deployment. Regardingadvanced techniques to localize users, the authors present how to make use oftemporal or spatial signal patterns, user collaboration, and motion sensors. Inthe second area of efficient system deployment, they discuss the advances on re-ducing offline labor-intensive measurement campaigns, techniques on adapting tofingerprint changes, calibrating heterogeneous devices for signal collection, andachieving energy efficiency for smart devices.

Fingerprinting is generally accepted to be the most promising and widelyimplemented technique, and for this reason the scope of this thesis is to mainlyconcentrate on this specific localization technique. As a method, it usually requiresan off-line and an on-line phase. During the off-line phase, the fingerprint radiomapis generated, either through RSS measurements or through simulations.

6


Table 1.3 – Indoor Positioning:Technologies comparison

Technology Advantages Disadvantages

Wi-Fi [2]Popular, good range, ac-curacy

Impractical until generating aradiomap. Requires severalAPs. Dynamic effects influ-ence performance

Bluetooth [28] Widely used Low range, Low accuracy

BLE [4]Low energy, range infor-mation is available

ow range, low coverage, easilyinterfered

RFID [29] NLOSSmall Range, limited cover-age

UWB [30]

High accuracy even inmultipath, limited inter-ference with other RF sys-tems

High cost, issues with metal-lic surfaces

VLC [5] High accuracyinfrastructure required, LOS,UE constraints

ZigBee [31] Low energy, Low Cost Easy to interfere

Infrared [32] LOS (secure coms)LOS, small range, vulnerableto lighting conditions

Ultrasonic[32]

does not interfere with RFLoss of signal due to obstruc-tions

IoT [9] Widely spreadLow range, sensitive to inter-ference

Dead Reckon-ing [33]

No infrastructure required losing accuracy over time

During the on-line phase, real time RSS measurements are performed by aMobile Station (MS) and processed through positioning algorithms, in order toestimate the location of the user.

In general, the positioning accuracy achieved by any fingerprint-based RTLS,is affected by various factors such as the number of APs, the utilized technology,the positioning algorithms, the device diversity, the complexity of the environmentand any dynamic radio effects. The degree that these factors affect the positioningaccuracy, is reflected initially on the quality of the fingerprint radiomap which isgenerated during the offline phase. Since the fingerprints populating the datasetare utilized in the positioning process, it can be safely claimed that each finger-print’s uniqueness -when compared to other fingerprints- defines the overall systemperformance. Hence, measuring and improving their uniqueness -or correlationlevel-, i.e. their quality, is of critical importance.Finally, it is also of great importance, for the scientific community and commercialkey players in the sector of localization, to objectively and reliably evaluate theperformance of an RTLS in order to provide valid results. Both aforementionedchallenges fall within the scope of this thesis as explained in Section 1.2.

7


Table 1.4 – Indoor Positioning: Typical Active Localization Systems

System TechnologyPositioning Al-gorithm

Accuracy

SpotON Active RFIDAd-Hoc Lateration(RSSI)

depends oncluster size

Ubisense UWBLeast Squares(TDOA, AOA)

99% within13cm

Ekahau RTLSWLAN -IEEE802.11a/b/g/n

Probabilistic KNN(RSSI Fingerprint-ing)

2− 3m

MicrosoftRADAR

WLAN KNN (RSSI) 24.3m

AeroScout WLANTOA Triangu-lation, TDOA,RSSI

1− 5m

Intel PlaceLab WLANMap-based triangu-lation

20− 30m

Skyhook WPSWLAN/Cellular/GPS

Hybrid99.8% within10m

PinPoint 3D-iDWLAN /UHF (40MHz) /Bluetooth

TDOA andBayesian approach

1m

MIT CricketBeacons/RF (418MHz)+Ultrasound

Least Square99% within30cm

1.2. Research Scope

A major challenge faced in all types of indoor positioning systems is to objec-tively evaluate the performance of different approaches and methodologies imple-mented, which combine heterogeneous technologies and fuse information from avast number of sources and sensors. This challenge has been the core motivationto set a twofold research scope in this thesis:

(1) To investigate the existing evaluation methodologies related to indoorfingerprinting positioning systems and platforms, identify possible weakareas and propose new techniques and algorithms that will contribute to-wards a more objective and unbiased performance evaluation of existing,or newly proposed indoor positioning methods and algorithms.

(2) To propose new indoor positioning algorithms with the aim to improvethe indoor localization accuracy achieved, by combining data from dif-ferent technologies.

Towards the evaluation of the indoor positioning methodologies, which is themain direction of this research work, we introduced 3 novel algorithms:

(1) Sample Size Determination Algorithm (SSDA) for the Evalua-tion of Fingerprint-Based Indoor Localization Systems: We pro-pose an algorithm that calculates the safest minimum sample size of po-sitioning data which is required for an objective performance evaluation

8

1.2. RESEARCH SCOPE

of fingerprint-based localization systems. The use of a correct, indepen-dent, unbiased and representative sample size can speed up the training,evaluation and calibration procedures of a fingerprint-based localizationsystem, while ensuring that the system’s true accuracy is achieved. Theproposed Sample Size Determination Algorithm (SSDA) takes into con-sideration the desired confidence level, the resulting standard deviationof a small sized preliminary sample, and the error approximation withrespect to the actual error of the system. Subsequently the algorithm pro-poses the final sample size for the evaluation and/or calibration and/ortraining of the utilized fingerprint radiomaps. Additionally, SSDA, as-sumes random sample allocation in the area of interest in order to avoidbiased results.

(2) Tolerance Based - Normal Probability Distribution (TBNPD)correlation algorithm for the quality evaluation of any fingerprint–based dataset: A novel algorithm is proposed for evaluating the qual-ity of RSS fingerprint–based databases, utilized by positioning systems.The proposed “Tolerance Based - Normal Probability Distribution” cor-relation algorithm, calculates the correlation level between each pair offingerprint entries forming the radiomap. Possible RSS fluctuations, oc-curring due to the dynamic nature of the environment, are also taken intoconsideration. The proposed TBNPD algorithm offers the possibility toassess the uniqueness of each fingerprint entry in a radiomap, prior itsutilization in a Real Time Positioning System. By applying this process,one has the opportunity to improve the positioning accuracy of an RTLS,by modifying the network infrastructure or its characteristics, in order toupgrade the quality of the initial input data i.e. each fingerprint unique-ness. In this way, different fingerprint radiomaps of the same buildingenvironment, can be compared and the best one can be chosen to feedthe positioning platform, leading to better accuracy.

(3) A Binomial Distribution Approach for the Evaluation of IndoorPositioning Systems: By adopting a binomial distribution approachduring the evaluation process, we introduced a practical approach thatallows the evaluation of the expected accuracy of an already deployedpositioning system, under its actual operating conditions. The data col-lected by the binomial experiment is statistically processed, in order forthe evaluator to asses the strength of his evidence, and finally decidewhether the declared accuracy can actually be met or not. The proposedapproach, due to its robustness, can be of significant value in a number ofscenarios including device diversity scenarios in dynamic environments.

With respect to the combination of data and technologies for improving indoorlocalization accuracy, several new methods were tested and proposed:

(1) Combining IEEE 802.15 (BLE) and IEEE 802.11 (Wi-Fi) technologies(2) Combining smart lighting and radio technologies(3) Utilizing Directional Antennas for improving a radiomaps’ quality char-

acteristics

In the same scope, further contributions were made in:

9


(1) Map-aided localization methods, where localization accuracy improve-ment was achieved by imposing map-constraints into the positioning al-gorithms in the form of a–priori knowledge.

(2) Single Access Point positioning, by exploiting polarization in MIMO de-vices.

(3) WSN optimization deployment and generation of radio–maps which canbe implemented in IoT and WSN localization, especially in indoor envi-ronments.

(4) Self-Organization Networks, where the generated radiomaps and the lo-cation of the mobile users are utilized towards reduced cost and energyper bit for future emerging radio technologies.

1.3. Structure of Thesis

The remainder of the work is organized as follows: Chapter 2 provides anintroduction to localization techniques and highlights the most dominant methodsused for location applications in indoor environment. Moreover, a literature reviewis performed, focusing on the scope of the main objectives of this work. Initially,the advances related with the fingerprint methodologies are presented, includingthe advantages and the inherent constraints that need to be addressed. Then, theexisting performance evaluation processes that are most commonly implemented inRTLS Systems are analyzed. Finally the chapter ends with an overview of hybridindoor positioning platforms that the research community proposed. In Chapter3, Chapter 4 and Chapter 5, our research contribution in the area of qualityevaluation is presented, supported by experimental performance evaluation, resultanalysis and scientific publications. In Chapter 6 we present several novel meth-ods that were developed, where different technologies are combined and data isfused, providing enhanced localization accuracy in indoor environments. Finally,Chapter 7 presents the conclusions and future research directions.

10

CHAPTER 2

Related Work

2.1. Introduction

This Chapter begins with a brief theory analysis of the basic localization tech-niques: Time–Based estimation, Direction–based estimation, Fingerprinting andDead Reckoning, as per section 2.2. A deeper analysis on the Fingerprint-basedmethodologies follows in section 2.3, since this approach is heavily utilized withinthe scope of the thesis, and all published algorithms and scientific work involvesexperimental testing based on this method. In section 2.4, a brief overview ofthe work behind information fusing for localization purposes, is performed. Thissection supports the second part of the thesis, which includes the presentationof novel localization methodologies, where information fusion from different radiotechnologies and non-radio parameters is proposed, in order to achieve increasedlocalization accuracy. Finally, in section 2.5 we discuss the different RTLS evalua-tion techniques and we identify the drawbacks and constraints of the performanceevaluation approaches, as they are typically described in the scientific papers. Theidentification of the research gaps in this area, was the motivation for the mainpart of this thesis, which led to proposing several algorithms in order to improvethe objectivity and reliability of the results.

2.2. Positioning Techniques

2.2.1. Time–Based Estimation

TOA is the one way propagation time of a signal travelling between the sourceand a receiver [19]. TOA measurements are used to determine the position of aMobile User (MU), based on the fact that the distance between the MU and thetransmitting source, is directly related with the time that a signal takes to travelthe aforesaid distance. During Time-based localization techniques implementation,several challenges need to be addressed [38]:

(1) An accurate estimation of the signal propagation delay is required, whichmeans that the existence of accurate clock synchronization between thenodes and the MU, is critical. Such synchronization can be avoided ifthe round-trip is measured. However, the measurement errors are alwayspart of the problem, especially due to the multiplication of time t with thespeed of light (c = 3x108(ms−1)) that will finally calculate the distancetravelled d = ct.

(2) Clock synchronization, as a task, is still costly and complicated.(3) Multipath effects and NLOS signals highly affect the accurate calculation

of signal propagation delay.

11

CHAPTER 2. RELATED WORK

A typical structure of a TOA positioning concept is shown in figure 2.1. As alateration problem, TOA positioning requires at least three transmitting sources(node, BS or AP) to localize a MU. Assuming that measurement errors are ab-sent, a circle is calculated around each transmitting source TXi, at a distance di,calculated based on the signal TOA. The intersection between all circles representthe possible location of the MU.

Figure 2.1 – Typical TOA Positioning Concept

Mathematically the model can be presented as follows [19]: We initially as-

sume that ~X = [x, y]T

is the unknown MU position, while ~Xl = [xl, yl]T

to be theknown coordinates of the lth sensor, where l = 1, 2, . . . , L and L > 3. The distancebetween the MU and the lth sensor, denoted by dl, is given by:

dl = ‖ ~X − ~Xl‖2 =

√(x− xl)2 + (y − yl)2 (2.1)

With no loss of generality, it is assumed that the source emits a signal at time0 and the lth sensor receives it at time tl; which means that tl are the TOAs whichhave the following relationship with the dl:

tl =dlc, l = 1, 2, . . . , L. (2.2)

12

2.2. POSITIONING TECHNIQUES

where c is the speed of light.In practice, the range measurements ~rTOA include an error denoted as ~nTOA.Hence, the range measurement is the sum of the actual distance dl as given byequation 2.1 rewritten in vector form, plus the error ~nTOA. Hence, vector wise theaforementioned concept is expressed by formulas 2.3 to 2.6.

~rTOA = ~fTOA(~ )x+ ~nTOA (2.3)

, where

~nTOA = [ ~nTOA,1, ~nTOA,2, . . . , ~nTOA,L]T

(2.4)

~rTOA = [ ~rTOA,1, ~rTOA,2, . . . , ~rTOA,L]T

(2.5)

and

~fTOA(~ )x =

√(x− x1)

2+ (y − y1)

2√(x− x2)

2+ (y − y2)

2

...√(x− xL)

2+ (y − yL)

2

(2.6)

At this point, it worth mentioning that ~nTOA,i are zero-mean uncorrelatedGaussian processes, with variances σ2

TOA,i which indicates LOS transmission [19].

Based on the above condition, the Probability Density Function (PDF) for eachrandom scalar variable RTOA,i is calculated by formula 2.7:

p(rTOA,i) =1√

2πσ2TOA,i

exp

[− 1

2σ2TOA,i

(rTOA,i − di)2]

(2.7)

A practical method that was relatively recently implemented in the TOA esti-mation, is the utilization if the Channel State Information (CSI). In such methods,the physical layer information processing, is utilized to estimate the TOA or bet-ter the round-trip time. Usually, these methods are constrained by the underlyingavailable bandwidth of the technology and require compatible hardware to esti-mate the TOA information [39], [40].

A more simple method that avoids the synchronization between the MU and thenodes is the TDOA. However, the synchronization between the nodes themselvesis still a prerequisite. TDOA is practically the difference in arrival times of theemitted signal, received at a pair of sensors as illustrated in figure 2.2.

Moreover, similar to the TOA, the exact locations of all base stations are con-sidered known to the localization infrastructure. Each calculated TDOA forms ahyperbola in the localization space and the intersection of all hyperbolas representsthe possible location of the mobile terminal. The TDOA measurement model canbe formulated as follows [19]:it has been assumed that the source emits a signalat an unknown time t0 and the lth sensor receives it at time tl, l = 1, 2, . . . , L with

L > 3. There are L(L−1)2 distinct TDOAs from all possible sensor pairs, denoted

by tk,l = (tk − t0)− (tl − t0) = tk − tl, where l = 1, 2, . . . , L with k > l. However,there are only (L-1) non-redundant TDOAs. For example if L = 3, the distinct

13


Figure 2.2 – Typical TDOA Positioning Concept

TDOAs are t2,1, t3,1 and t3,2. However, t3,2 = t3,1 − t2,1, which is redundant. Thesolution of TDOA approach is similar to the TOA as described previously.

Summarizing, the Time–based estimation, is a promising but complex local-ization technique, requiring high accuracy clocks for station synchronization, or atleast accurate time measurement.

2.2.2. Direction–Based Estimation

Direction of Arrival (DOA) is the arrival angle of the emitted signal observedat the receiver [19]. Consequently, DOA techniques can be utilized in localization,by determining the direction of propagation of a signal and its angle of arrival ata minimum of two nodes as shown in figure 2.3.

The mathematical model of DOA method is based on the geometrical formulaof angle φl, which represents the DOA between the source and the lth receiver, orin other words, the angle between the Line of Bearing (LoB) and the x-axis, in2D-space:

tan(φl) =y − ylx− xl

, l = 1, 2, . . . , L (2.8)

with L > 2.

14


Figure 2.3 – Typical DOA Positioning Concept

As in the case of TOA, during actual DOA measurements denoted as rDOA,l,also an error nDOA,l is included, which occurs from the noises -assumed to bezero-mean uncorrelated Gaussian processes with variances σ2

DOA,l. Hence the cal-culation of rDOA,l is given by the formula 2.9:

rDOA,l = φl + nDOA,l = tan−1[y − ylx− xl

]+ nDOA,l, l = 1, 2, . . . , L (2.9)

Vector wise, 2.9 can be re-written as:

~rDOA = ~fDOA(~ )x+ ~nDOA (2.10)

where

~nDOA = [ ~nDOA,1, ~nDOA,2, . . . , ~nDOA,L]T

(2.11)

~rDOA = [ ~rDOA,1, ~rDOA,2, . . . , ~rDOA,L]T

(2.12)

and

~fDOA(~ )x = ~φ =

tan−1

(y−y1

x−x1

)tan−1

(y−y2

x−x2

)...

tan−1(

y−yLx−xL

)

(2.13)

15


Similar to other localization techniques, various DOA estimation methodolo-gies have been proposed. Typical examples include a spatial spectral estimationmethod, called Delay and Sum (DAS) [41] as well as eigenstructure methods, suchas Multiple Signal Classification (MUSIC) [42], root MUSIC [43], and Estimationof Signal Parameters via Rotational Invariance technique (ESPRIT) [44]. Recentproposed approaches, taking advantage of the technological evolution, introducehardware novelties, such as the usage of Software Defined Radios (SDRs) and/orpropose the fusion of multiple techniques in order to combine the benefits of eachcategory. In this way, researchers in [41] implemented the localization techniqueon an SDR and fused DAS and root-MUSIC algorithms. More specifically, theytook advantage of the simplicity of DAS which is suitable for on-line, high speeddetection, with less power consumption, while for off-line DOA estimation theyimplemented the more sophisticated root-MUSIC. Finally, in more recent work,authors of [45], proposed the utilization of sectorized antennas instead of antennaarrays. They introduced a DOA estimator that is broadly applicable with differ-ent sectorized antenna types and signal wave-forms, and has low computationalcomplexity.

Summarizing, the Direction–based estimation has been proposed for localiza-tion several decades ago. It does not require clock synchronization but still aseries of constraints exist, such as the necessity of sophisticated antenna arraysand computational complexities. For this reason, as a technique is not preferableor attractive when it comes to broad commercial applications.

2.2.3. Received Signal Strength Estimation

Received signal strength estimation, is based on the calculation of Path Lossfrom each transmitting antenna (Base station or Access Point) to the receiver po-sition. By accurately calculating the path loss, we can estimate the radio coverageof each transmitter, as well as the signal strength expected to be received by eachuser equipment (UE), within the area of the radio coverage. Path loss is definedas the ratio of the received power Pr to the transmitted power Pt and typically isexpressed in dB [46] as shown in equation 2.14.

PL(dB) = 10 log(Pr

Pt) (2.14)

In the most simple, free space theoretical scenario, where the propagationrefers to Line of Sight (LOS) between the transmitter and the receiver, the pathloss follows an inverse square law with respect to the distance R. It also dependson the frequency -aka the wavelength λ - and the antenna patterns of both thetransmitter and the receiver which specify the gains Gt and Gr respectively as perequation: 2.15.

PLfree(dB) = 10 log

(GtGrλ

2

(4π)2R2

)(2.15)

For Gt and Gr equal to unity [47], the path loss equation 2.15 can be simplifiedto

16


PLfree(dB) = 32.4 + 20 logR(km) + 20 log f(MHz) (2.16)

The free space path loss model, lucks practical implementation, and for thisreason several empirical and semi-empirical models were developed, mostly rep-resenting specific categories of outdoor environments[48]. They are usually ex-pressed by a set of simple, easy to implement formulas, which are extracted basedon a series of measurements in the environment that they represent and for thefrequency range that they are applicable. Typical examples of such models are:

(1) Okumura-Hata Model, which is fully empirical and is based on exten-sive measurements carried out in Tokyo in the frequency range 150MHz−1.5GHz [49]. Estimation of the path loss, is performed through a seriesof graphs, as analyzed in [50].

(2) Cost-231-Hata model that is an extension of the Okumura-Hata, andcovers the frequencies 1500MHz − 2000MHz [51].

(3) Ibrahim and Parsons Model that attempts to quantify urban prop-agation loss around London [52].

(4) Allsebrook and Parsons Model, which is a semi-empirical model forsuburban predictions and it is based on measurements made in Britishcities, at 86MHz, 167MHz and 441MHZ [53].

(5) Walfish-Bertoni Model, also a semi-deterministic model that predictsmultiple diffraction over building rooftops [54].

All the above models can provide a relevant draft estimation of the path lossin an outdoor environment, but they cannot perform reliably and accurately eitherin complex outdoor environments.

Several studies have been performed in an effort to develop empirical models thatcan capture the radio waves behaviour in complex indoor environments [55], [56],[57]. Based on the aforementioned studies, when considering the large-scale atten-uation model, most researchers model the radio propagation path loss as a functionof the attenuation exponent n as indicated in equation 2.17. In this equation, Pd

is the power at distance d from the transmitter and Pd0is the power at a reference

distance of 1m. n exponent has a value of 2 for free space cases and increases forindoor environments. In this concept, authors of [58] defined a value of 2.6 fortheir specific site and managed to achieve an average difference of 3.0dB betweenmeasurements and estimates.

Pd[dB] = Pd0[dB]− 10n log

(d

d0

)(2.17)

In cases where accuracy is important, deterministic modeling is the only vi-able solution. Deterministic or Site-specific Models are based on the applicationof electromagnetic techniques and numerical methods (such as Ray Tracing) to asite-specific environmental description [59]. The environment is described in termsof the position of the objects, their dimensions, orientation and material consti-tutive parameters. The wireless network topology and configuration parametersare also required. Ray Tracing is the dominant deterministic technique used topredict propagation effects in mobile and personal communication environments.It is based on Geometrical Optics (GO) and Uniform Theory of Diffraction UTD;

17


an approximate method for predicting a high-frequency electromagnetic field. RayTracing identifies all possible ray paths between the transmitters and the receiverlocations and implements high-frequency electromagnetic theory to calculate am-plitude, phase, time delay and polarisation parameters.

2.2.4. Fingerprinting

Fingerprinting is the most common and widespread localization technique, es-pecially when it comes to indoor environments[37], although it can be also appliedin outdoor environments [24]. This technique is typically based on measuring sev-eral radio (usually the RSSI) and/or non-radio parameters and linking their valueswith each location in the area of interest. This type of profiling is stored in a data-base and retrieved later by the mobile user, during the position estimation request.At this point the localization algorithm performs a matching procedure betweenthe dataset fingerprints and the real time measurements and calculates the userposition. [60].

More specifically, fingerprint-based positioning is separated in two phases. Dur-ing the first, off-line phase, also known as the radiomap generation, a database isconstructed. Each set of data (fingerprint) relates a specific position in the areaof interest with a number of signal parameters such as the Cell Identifiers (CIDs)of the discoverable APs, the RSS values of each AP etc. These fingerprints arestored in the form of vectors.

Figure 2.4 – Fingerprinting: Offline Phase

18


It is understood from the above that, in order to create the radiomap, ablueprint of the study area needs to be utilized and the measurement locations tobe determined as indicated in figure 2.4.

Figure 2.5 – Fingerprinting: Online Phase

Then a measurement campaign is launched and each data vector is directly re-lated with the coordinates of the position where the observation took place withinthe study area. In order to elaborate with any possible changes in RSS, which mayoccur due to different UE antenna orientations, a separate set of measurements isrecorded for each device orientation. The aforementioned RSS differences are anunavoidable inherent problem of the non-isotropic antenna patterns of the devices.Further to the above, a number of measurement samples is also taken into consid-eration and the mean RSS value is typically utilized, in an effort to manage thetime variation of the RSS. Additionally, the coordinates of the APs -if known- canbe considered as reference points that can be utilized by the positioning algorithmsin order to improve positioning accuracy.

The offline phase can be performed either by utilizing software site survey toolssuch as inSSIDer [61], sophisticated measurement equipment (spectrum analyz-ers), or it can be generated through simulations from full 3D Ray tracing toolssuch as TruNET wireless [62]. More enhanced fingerprinting positioning systems,may include additional radio and/or not radio parameters, such as information

19


from inertial sensors, to improve the uniqueness of each fingerprint [63].

The second, on-line phase, of the positioning procedure, refers to the associa-tion of the observed RSS values of the Mobile Station (MS), with the generatedradiomap fingerprint entries and the location estimation procedure. As illustratedin figure 2.5, based on the available signal parameters observed by the MS atthe unknown position, several deterministic and/or probabilistic positioning algo-rithms are implemented, aiming to achieve the best match and retrieve an estimateof the position.

Some of the common positioning algorithms include K-Nearest Neighbour (KNN),Weighted K-Nearest Neighbour (WKNN), Minimum Mean Square Error (MMSE)and Maximum A Posteriori (MAP). Details can be found in [64],[65] and [66] andsome brief analysis in subsection 2.3.3.

Since Fingerprint–based localization is on of the most popular and applicablelocalization techniques for indoor environments, the scope of this thesis is heav-ily focused on the aspects of this technique. Detailed analysis on the utilizedfingerprint-based methodologies is provided in section 2.3.

2.2.5. Dead Reckoning

Dead reckoning is the technique of calculating the mobile user’s position byusing a previously determined position, taking into consideration several parame-ters, such as speed, direction etc. Systems implementing dead reckoning typicallyuse sensors attached on the user able to estimate relative, rather than absolutelocation i.e. the change in position since the last update, by knowing angle θi in[°] and ri in [m] as shown in figure 2.6.

The main advantage of this technique is that infrastructure requirements arenil or minimal. On the other hand, the disadvantage is that the localization erroraccumulates with time. For this reason it is usually combined with other localiza-tion techniques and work supportively, exchanging and retrieving information forcorrecting the relative location [33].

With the sensors and processing nodes getting smaller and more affordable, sev-eral solutions have been proposed for Pedestrian Dead Reckoning (PDR) systemsespecially for indoor tracking. These solutions use inertial and other sensors, of-ten combined with domain-specific knowledge about walking, in order to trackuser movements. There is currently rich research and relevant literature, which issummarized in [33].

In [67], authors presented a smartphone-based pedestrian dead reckoningmethod named SmartPDR, which tracks pedestrians through typical dead reck-oning approach, using data from inertial sensors embedded in off-the-shelf smart-phones, without the support from any other infrastructure or system.

In another approach in [68], researchers fused data retrieved from dead reck-oning sensors with sound–based relative localization, using the extended Kalman

20

2.3. FINGERPRINT-BASED POSITIONING METHODOLOGIES

Figure 2.6 – Dead Reckoning Concept [source: PNI Sensor Corp. (Santa Rosa,Calif.)]

filter (EKF) algorithm, in a multi-robot distributed cooperative localization method.In this way, the authors aim to correct the accumulative error of dead reckoning.

Finally, in [69], authors proposed an effective indoor localization method forautomated guided vehicles (AGVs) utilizing dead reckoning but eliminating its in-herent drawbacks, by utilizing an encoder, and particle filters with a 2D laser rangefinder. The proposed method involves an asynchronous localization algorithm.

2.3. Fingerprint-based Positioning Methodologies

As described previously in subsection 2.2.4, fingerprint-based positioning isseparated in two phases; the offline and the online phase, as illustrated in figure2.7.

Focusing on the offline phase, two typical methods exist for the collection ofthe necessary radio parameters; either with measurements or through simulations.

2.3.1. Measurements

A large number of data samples are collected at several predefined positionsin the area of interest, using specialized calibrated equipment and software. Theconditions for the data sampling at each point, such as the antenna height andorientation, should be maintained constant in order to ensure consistent finger-prints. Measurement campaigns are generally costly and time-consuming, whilethe generated radiomaps provide a static perception of the area of interest. Suchradiomaps are hard to maintain in case of indoor area structural changes (addi-tion of new furniture or renovation works), or modifications in the network setup(addition of new APs, changes in antenna orientation or transmitter power etc.)

21


Figure 2.7 – Fingerprint-Based Localization Concept

Examples of utilization of the measurement method can be found in [70], [65] and[71].

22


2.3.2. Theoretical Estimation using Radio Propagation Modelling Tech-niques

An alternative radiomap generation method, that gained ground in the recentyears, due to the evolution of computer calculation performance, is the utiliza-tion of radio planning simulation tools. Statistical, semi-deterministic or fullydeterministic propagation models are used for the theoretical estimation of the ra-diomap fingerprints. The Wall Attenuation Factor (WAF) model which is basicallyan extension of the free space path-loss model with an addition of the attenuationcaused by walls, was implemented in [70] and further evaluated in [72]. Empiricaland semi-deterministic models such as One-Slope Model, Multi-wall Model andMotif Model, are examined in [73], while other Localization Statistical Models areproposed in [74] and [75]. The 2D Ray Tracing models adopted in [76] and [77]include the reflections from the various indoor environment structures, while full3D Ray Tracing Models utilized in [78], [79] and [80], take into account the full 3Denvironment dimensions. Overall, compared to measurements, radio simulationsare much faster, less expensive to implement and one can easily re-run them incase the environment changes [79].

2.3.3. Fingerprinting Positioning Algorithms

During the on-line phase, different positioning algorithms are used for match-ing the observed measurements with the offline fingerprints in order to estimatethe user position. Fingerprint–based positioning using RSS can be classified intotwo main categories (1) Deterministic and (2) Probabilistic approaches. The de-terministic positioning methods estimate location as a convex combination of thereference locations [81]. Usually, the K reference locations with the shortest dis-tance between ri and s in the n-dimensional RSS space are used and the estimated

location is given by equation 2.18.

=

K∑i=1

(wi∑Kj=1 wj

`′i

). (2.18)

The set {`′1, . . . , `′l} denotes the ordering of reference locations with respect toincreasing distance between the respective fingerprint ri and the observed mea-surement during positioning s, i.e. ‖ri − s‖. The distance can be calculated usingstandard norms, such as the Manhattan (1-norm) [64], the Euclidean (2-norm)[70], the general p-norm with modifications [71] or the Mahalanobis norm thatemploys the sample means and variances of the reference fingerprint [82].

One possible option for the non–negative weights wi in Eq. (2.18) is the in-verse of ‖ri − s‖ and in this case the positioning method is known as WeightedK-Nearest Neighbour (WKNN) [64]. The K-Nearest Neighbour (KNN) methodassumes equal weights for the candidate reference locations, while setting K = 1leads to the simple Nearest Neighbour (NN) method [70, 83, 84]. In generalthe KNN and the WKNN methods have been reported to provide higher level ofaccuracy compared to the NN method, particularly with parameter values K = 3and K = 4 [70, 64]. However, if the density of the RSS radio map is high, NN

23


method performs equally well compared to more complicated methods [81]. Sev-eral variants of the KNN method have been discussed in the literature, includingthe Database Correlation Method (DCM) which introduces an additional term inthe error function to penalize missing RSS values in the fingerprints [85, 86].

In probabilistic methods, location ` can be estimated by calculating and maximiz-ing the conditional posterior probabilities p(ì|s), i = 1, . . . , l given an observedfingerprint s and a fingerprint database (l is the number of fingerprints in thedatabase). These methods have been extensively used in the Maximum A Posteri-ori (MAP) approach [87, 88, 89] and the Minimum Mean Square Error (MMSE)approach [65] to estimate the expected value of `.

The posterior probability p(ì|s) is obtained by applying Bayes’ rule, as indi-cated in equation 2.19.

p(ì|s) =p(s|ì)p(ì)∑li=1 p(s|ì)p(ì)

(2.19)

where p(s|ì) is a conditional probability calculated through statistics at thesurvey stage and p(ì) is the a–priori probability, a weighting factor based onthe probability distribution of the target over the reference position candidates(database fingerprints). If we assume that we do not have any prior knowledgethen this prior can be assumed to be unity providing equal a-priori probability toall the fingerprint candidates in the database.

2.3.4. Handling the constraints of fingerprinting

As a technique, fingerprinting, can be implemented on any available technol-ogy or combination of technologies. Several approaches were also proposed by theresearch community, in an effort to handle the core constraints and difficulties ofthe technique: the labor-hungry procedure of the offline phase, the difficulties tokeep the fingerprint dataset up-to-date, the influences of the dynamic nature ofthe environment and the inherent problems of each technology that is used everytime. Authors in [90] presented the base-station-strict (BS-strict) methodology,which emphasizes the effect of BS identities in the classical fingerprinting. Theyalso included particle filters in road network information to improve accuracy.Their approach was tested in the city of Brussels utilizing a WiMAX network.The computational complexity of the online localization phase is discussed in [91],where researchers proposed clustering of the collected fingerprints, with the useof corrected K-Means algorithm, in order to achieve a more simplified and lightlocalization algorithm.

A more challenging step, towards 3-dimensional fingerprint localization, basedon ZigBee standard, is presented in [92]. In another approach, for providingcontinuous and easy generation of fingerprints, authors of [93], proposed an in-door sub-area localization scheme, based on fingerprint passive crowd-sourcingand unsupervised clustering. The suggested method, first classifies unlabeled RSSmeasurements into several clusters and then relates clusters to indoor sub-areas

24


to generate sub-area fingerprints.

To handle the localization problem of multipath propagation in complex indoorenvironments, a fingerprint based localization method of signal subspace match-ing, is proposed in [63]. In this method a small antenna array is used in an indoorenvironment. It is claimed that, compared to the RSSI, the signal subspace finger-print can provide better results by utilizing more space information. The signalsubspace fingerprint is created from the processing of the received signal with self-correlation and decomposition of it’s eigenvalue. Location is then determined byfinding the smallest angle between the online received signal subspace and the onesstored in the relevant fingerprint database.

In [94], researchers consider the resource limitations of wireless networks and pro-pose an improved fingerprint-based localization approach that adapts a path lossmodel for fingerprint creation and localization. Based on the proposed approach,they present two related localization schemes. The first is a path-loss-based fin-gerprint localization (PFL) scheme and the second is a dual-scanned fingerprintlocalization (DFL) scheme. The PFL attempts to improve positioning precision,and the DFL attempts to guarantee positioning reliability. Several simulationswere performed, indicating that the proposed schemes improve the positioningprecision and reliability in resource-limited environments.

Furthermore in [95], researchers take advantage of deep learning, extreme learningmachines, and high level extracted features by an auto-encoder to propose an algo-rithm that seems to improve the localization performance in feature extraction andclassification. Additionally, the number of training data is increased so that thesystem adapts to the dynamic nature of the environment and gradually improveslocalization performance. Working in the same concept of automatically updatingthe training dataset of the fingerprint database, authors in [96] propose an Ag-ing Resistance Wi-Fi Fingerprint-based Localization (AR-WFL) system based oncrowd-sourcing. The system includes an update phase that can reflect environmen-tal changes periodically and prevent performance degradation. In addition, theyadopt active learning scheme with an uncertainty selective sampling algorithm tomaximize cost-efficiency of the update phase. They evaluated the performanceduring update phase in terms of location estimation accuracy using a dataset theycollected for 2 months with the results indicating a slight improvement comparedwith a system without an update phase.

For the rapid generation of the fingerprint database, authors of [97], proposed theuse of a mobile data-collection cart, equipped with a laser rangefinder and a smart-phone. In practice, as some regions might be inaccessible, the researchers adopteda kriging-based interpolation method that exploits the spatial auto–correlation ofthe RSSI to efficiently generate and update the fingerprint database. To over-come the instability of RSSI and improve positioning accuracy, they proposed atime-variant multi-phase fingerprint map. In their approach, numerous fingerprintdatabases are constructed for different time periods, which are then automaticallyselected with respect to the most appropriate one, according to the time period.

25


In [98], an artificial neural network learning algorithm is implemented togetherwith the conventional fingerprinting method. More specifically, a parallel learn-ing method is proposed, in an effort to reduce the error introduced by the indoorheight and an indoor positioning data augmentation method for data generaliza-tion.

Finally, research work has been performed towards improving the quality of fin-gerprints collected during the offline phase. A part of the problem occurs whensome abnormal RSSI values in the fingerprint database, influence the localization/ matching procedure, resulting in low precision. In order to decrease the influ-ence of original data containing bizarre RSSI values, a filtering algorithm basedon the neighbor mean filter (NMF) is applied in [99]. The concept of NMF lieson replacing each pixel value in an image with the mean value of its neighbors,excluding itself. Adopting the NMF approach in RSSI fingerprint databases, theauthors attempt to identify the problematic RSSI values and replace them withthe mean nearby values. Simulations indicate that the optimized RSSI fingerprintdatabase can increase the accuracy of indoor localization.

Towards the same goal of developing more accurate fingerprint databases, in [6],we proposed the use of directional antennas. The aim is to position and orient adirectional antenna in such a way, as to reduce the correlation level of the RSSfingerprints forming the radiomap, while at the same time not degrading the cov-erage of the wireless network. Our work, among other novelties, is presented inchapter 6, section 6.4.

A list of typical localization platforms utilizing the IEEE 802.11 standard forfingerprinting purposes ([37]), is presenter later in Chapter 6, in table 6.2. Specialnotice is given to φmap, which is an experimental localization platform developedin a modular and expandable architecture in order to support the research workof this thesis.

2.4. Hybrid indoor RTLS and sensor information fusion

In the era of Internet of Things (IoT) and Body Sensor Networks (BSN),with the evolution and coexistence of other wireless technologies, several scenariosexploit information fusion in order to provide more accurate and reliable indoorlocalization. User position is more than ever critically linked with the availabilityof a wide range of services based on the user behaviour ([100], [101]). Typi-cal scenarios enabled by BSN technologies, include m-Health, e-Sport, e-Fitness,and e-Wellness. In all these applications, numerous programmable wireless sen-sors, with enhanced capabilities, are combined and configured in order to directlymonitor several parameters, in a non-invasive way [102]. A comprehensive andsystematic review of the state-of-the-art techniques on multi-sensor fusion in theBSN research area is provided in [103].

In this section, we will discuss the most typical approaches mentioned in the

26

2.4. HYBRID INDOOR RTLS AND SENSOR INFORMATION FUSION

literature that combine and fuse information retrieved from popular wireless tech-nologies in order to support the localization demands of IoT and BSN applications.

A well investigated approach for indoor localization, is the combination of the twomost widely spread technologies: the Wi-Fi and the Bluetooth or the relativelyrecently introduced BLE (Bluetooth Low Energy). BLE was introduced as part ofBluetooth 4.0 specifications, allowing the devices to support both BLE and clas-sic Bluetooth protocols simultaneously [104]. The power efficiency of Bluetoothwith low energy functionality was especially created for IoT applications. It allowsdevices to run for long periods on extremely low power sources, such as coin-cellbatteries or energy-harvesting devices.

In the same area, [105], the authors proposed the creation of study groups infuture smart libraries, featuring a smart phone application to create study groupsand a hybrid BLE and Wi-Fi indoor positioning system. Their hybrid indoorpositioning system calculates two probable user locations every time. They areutilizing each technology separately, compare the estimated conditional probabil-ities and select the most reliable. Following a different approach, researchers in[106] assume a very dense IoT environment with BLE compatible devices and pro-pose an Iterative Weighted KNN (IW-KNN) indoor localization method. Theirmethod is based on RSS measured by the BLE devices which, due to a low powerconsumption, they have long life expectancy. Finally, a combination of Wi-Fi andBLE fingerprints was implemented in [107], utilizing the conventional WKNN al-gorithm. The test case included the deployment of 17 Estimote BLEs on top of anexisting Wi-Fi network, in a study area of 52m x 43m. Positioning was performedby utilizing both BLE and Wi-FI RSS fingerprints, resulting in a 23% accuracyimprovement of the RTLS system.

Working within the general concept of information fusion, authors in [108], pro-posed a combination of Wi-Fi and Vision integrated fingerprint (dubbed Wi-ViFingerprint) trying to achieve more accurate and robust indoor localization. Themethod consists of two steps: fingerprint mapping and fingerprint localization.In the mapping step, the Wi-Vi fingerprints, for all the sampling positions, arecomputed by using EXIT signs as landmarks. During the online phase, the ap-proach includes coarse localization with Wi-Fi matching, image-level localizationwith visual matching, and finally metric localization for localization refinement.The proposed method has been tested in an indoor office building of 11,200 m2with different types of smart-phones. The final localization error after the metriclocalization is claimed to be less than half meter in average.

In [109], inLoc is introduced, which is a positioning and tracking system usingcommercial mobile devices, with a navigation (routing) capability. With inLoc,authors propose, among others, a method for independent fusion of location in-formation from phone Inertial Measurement Unit (IMU) sensors and BLE beacons.

Summarizing, the information fusion for the purposes of enhancing indoor lo-calization accuracy, is expanding rapidly and is one of the most research active

27


topics. This thesis provides significant contribution, through the combination ofWi-Fi with BLE and Wi-Fi with Visual Light Positioning (VLP) technologies inindoor fingerprint-based localization. In [4], general localization data is gatheredfrom BLE devices and used to optimize the IEEE 802.11 RSS fingerprint-basedindoor positioning system. Implementing this concept, a new enhanced KNN po-sitioning algorithm was developed (i − KNN), which is able to filter the initialWi-Fi fingerprint dataset (radiomap), taking into consideration the proximity ofthe RSS fingerprints to the BLE devices. By choosing to filter the initial datasetinstead of simply combining BLE and Wi-Fi fingerprints, i − KNN utilizes anoptimized small size subset of possible user locations for the final position estima-tion. The proposed methodology achieves fast positioning estimation due to theutilization of a fragment of the initial fingerprint dataset, while at the same timeimproves positioning accuracy by minimizing estimation errors. In [5] a similarapproach is used but instead of utilizing BLE, the VLP system is adopted into thepositioning algorithm. Both approaches are analyzed in Chapter 6 (sections 6.2and 6.3 respectively).

2.5. RTLS Performance Evaluation Techniques

The variety of the existing commercial and experimental localization plat-forms, in combination with the diversified approaches they implement for user/assetposition estimation, makes any evaluation attempt to be extremely challenging.The research work in this area will be analyzed in this chapter and is stronglyinterconnected also with the next section 2.6, where the optimum sample selectioncriteria are discussed, for the purpose of objective and reliable evaluation of RTLS.

Initially, the evaluation trend was to develop benchmark standards which couldbe used to compare the performance of different schemes in a transparent andreproducible manner. Authors in [110] followed this approach and proposed theinclusion of different categories in a benchmark. They suggested mainly a cat-egorization in terms of the environment type (e.g. airport, train station), itsrelative dynamics (e.g. heavy movement of machines or people) and the dynamicsof the Mobile Station (MS). The authors analyzed the aforementioned categoriesin further critical factors, noticing among others the necessity of a detailed floorplan, the qualitative description of the different types of building materials, thetype, the configuration and the antennas of the APs and the MS. They also high-lighted the importance of the sample data, since such data is used to computethe position estimates. In this respect, they proposed the adoption of parametersextensively used, such as the cumulative distribution function,the median errorand multiples of the standard deviation, during the evaluation phase. Althoughthe bench-marking aims to produce comparable results, it is still a complicatedand abstract procedure to follow. At the same time there is no prioritization ofthe large number of factors taken into account. Following a relatively similar, butslightly more practical direction, CRAWDAD [111], a Community Resource forArchiving Wireless Data at Dartmouth, hosts a number of wireless datasets builton real users behaviour, with an aim to help researchers to practically identify andevaluate a number of problems. In this scope, [112] use a 2.4 GHz link measure-ment data set obtained from CRAWDAD, to evaluate three location distinction

28

2.6. THE CHALLENGE OF SELECTING THE RIGHT TESTING SAMPLE

methods.

Later, researchers in [113] and [114] enumerated a number of critical factors thatinfluence the performance of localization platforms, but without directly examin-ing the impact and relation to each other. Such factors included the number ofthe APs and their geometrical deployment, the furniture and internal partitionsetup, the humidity levels etc. The difficulty to assess the performance of differ-ent localization systems due to different testing conditions, was also reported in[115]. In the aforementioned research work, authors examined a set of localizationmechanisms and evaluated their performance under various configuration settingsand dynamics, using two building environments, one regular office building andone underground floor-plan. More recently, an attempt to assess explicitly theimpact of device diversity was performed in [116], [117] and [80]. In [117] and[80], authors combined linear transformation with a wide Kernel estimation tofurther reduce the positioning error, while in [116] performed measurements inan anechoic chamber and calculated the signal strength offsets between differentdevices, in order to calibrate radiomaps from different devices. Finally, on a differ-ent approach, Kjærgaard in [118] introduced Hyperbolic Location Fingerprinting(HLF) to address signal strength heterogeneity.

Summing up, it can be concluded that RTLS performance evaluation is extremelychallenging, due to numerous interconnected factors that affect the outcome. Itseems that there is a missing gap of standardized procedures and algorithms thatcan help and guide the research community to extract objective, reliable and mostof all comparable conclusions. Towards this direction this thesis contributes withthe work presented in chapters 3, 4 and 5.

2.6. The Challenge of Selecting the Right Testing Sample

In order to evaluate, calibrate, train or test a Fingerprint-based RTLS we needto retrieving a measurement sample. The importance of sampling is highlightedin [110], where the authors proposed the development of a benchmark standard.They specifically state that samples constitute the core of any benchmark forlocation systems, since they are used to compute the position estimates. Theyproposed that the benchmark specification should state the number of samplesrecorded per second and the duration of the measurements per location.

When trying to review common practices, literature suggests that researcherstend to utilize a diverse number of sample sizes for the purpose of evaluatingtheir research work, without necessarily clarifying the rationale behind the sampleselection. In [65], 40 observations were recorded for a set of 155 calibration points,that were used as training data to eliminate the randomness of human behaviour.In [66], authors measured one sample per second, for a period of five minutes (300samples total), while trying to investigate wireless channel changes over time.

Authors in [119] proposed a dynamic hybrid projection (DHP) technique forimproved 802.11 localization. During their experiments they collected 802.11 RSSdata at 27 different reference locations in the area of interest, on different days and

29


at four different user orientations. Out of this sample they selected 15 locationswith a step of 1.5− 2m, which they then used as training data.

A different sample size was used in [120], where different filtering strategiesfor real life indoor 802.11 positioning systems were analyzed and compared. Theauthors measured the radio distribution at 250 uniformly distributed grid pointsin an area of 15m x 35m.

In [121], the differences among the received signal strengths from a numberof 802.11 adapters were investigated. The authors of this work conducted systemvalidation using a total of 3120 positioning requests.

In another aspect of indoor positioning, authors of [122] introduced severalfault models to capture the effect of failures in the wireless infrastructure. Duringthe investigation of fault tolerance of positioning methods and evaluation in termsof their performance degradation, they contacted experiments based on a radiomapconsisting of 107 distinct reference locations having a step of 2 − 3m. A total of3210 reference fingerprints, corresponding to 30 fingerprints per reference location,were collected at the rate of 1 sample/sec. For testing purposes they collectedfingerprints along a path, consisting of 192 locations. Authors of [123] proposed anovel indoor localization scheme based on sub-area fingerprint determination andsurface fitting. During the performance evaluation of the proposed technique, theyperformed experiments in an area 16.2m x 28.5m by setting 25 reference pointsper room and randomly selecting 200 test points in the environment.

Finally, authors of [124] and [125] worked towards the challenge of deploy-ment load reduction in RSS based indoor localization systems. In their work,they proposed an interesting scheme that combines the data retrieved from a raytracing simulator with a limited number of measurements (15 % - 30% of thecomplete fingerprint dataset) and performs localization using manifold alignment.The aforementioned methodology leads to a significant load reduction but assumesthat the utilized fingerprint datasets have stronger correlation among neighbour-ing data points compared to other points. This assumption is not always valid inindoor environments with strong multipath effects in rayleigh channels.

From the above, it is observed that researchers tend to calibrate, train, test andevaluate indoor positioning systems by utilizing a variety of sample sizes and sam-ple patterns. One of the tasks of this thesis was to introduce a generic methodologyto calculate a sample size that will capture a set of predefined accuracy criteria,within the desired confidence level, for any test environment. This methodologyis expected to contribute towards the standardization of evaluation procedures forindoor fingerprint-based positioning systems.

Concluding, it is noted that within the research scope of this thesis, and towardscontributing in the improvement of evaluation methodologies, three novel algo-rithms were introduced. In [1] we propose Sample Size Determination Algorithm(SSDA), that calculates the safest minimum sample size of positioning data whichis required for an objective performance evaluation of fingerprint–based localiza-tion systems. Detailed analysis and performance testing is presented in chapter3. In [2] we present Tolerance Based - Normal Probability Distribution (TBNPD)correlation algorithm for evaluating the quality of Received Signal Strength (RSS)

30

2.6. THE CHALLENGE OF SELECTING THE RIGHT TESTING SAMPLE

fingerprint databases, before they are utilized by positioning systems. The al-gorithm and the evaluation software tool developed are explained in chapter 4.Finally, in [3] a more practical methodology was introduced, for the evaluation ofalready deployed positioning platforms, based on the binomial distribution. Chap-ter 5 elaborates on the details of this method.

31

CHAPTER 3

Sample Size Determination for the Evaluation ofIndoor Localization Systems

3.1. Introduction

As described in detail in Chapter 2, fingerprint-based positioning requires thegeneration of a dataset of measurements, usually the RSS, during an off-line phase.This dataset, called radiomap, requires calibration before being utilized for theestimation of the user location during the on-line phase. Calibration is impor-tant in order to train and configure the positioning algorithms to perform betterfor the specific environment ([79]). The calibration techniques use a sample ofmeasurements, which is taken in the area of the indoor environment. The samplesize, as well as the allocation of the samples in the area of interest, influence theoverall accuracy of the localization system. A similar sample of measurements isalso used for the performance evaluation of the system. The calibration and eval-uation procedures are two important steps that are influenced by the quality ofthe sample measurements. Selecting a small sample size or a biased sample can re-sult in misleading calibration parameters and a degraded accuracy during locationestimation. On the other hand, large sample sizes are more time consuming andmore expensive to carry out. To the author’s best knowledge, no previous workexisted for selecting the aforementioned sample sizes in fingerprint-based indoorlocalization systems. The main goal of the work presented in [1] was to developand suggest an algorithm that will define the minimum sample size which willensure correct calibration or training of the RTLS, and objective evaluation of thesystem’s performance.

3.2. Proposed Approach

In [1], the Sample Size Determination Algorithm (SSDA) is proposed, forcalculating the minimum, independent, unbiased, representative sample size ofpositioning data, as this applies to fingerprint-based indoor localization systems.Minimum in the sense that, it is large enough to ensure that, the estimated meanpositioning error of the system, lies within the desired confidence level, hence theevaluation and/or calibration results are within the set reliability criteria. Inde-pendent, unbiased and representative in the sense that the sample positions need tobe selected randomly within the area of interest, in order to avoid convenient pat-terns or specific samples that may cause biased results which can be questioned.The first step of the methodology requires the selection of a small preliminarysample of size nsps, the identification of the desired confidence interval ci and the

33

CHAPTER 3. SAMPLE SIZE DETERMINATION FOR THE EVALUATION OFINDOOR LOCALIZATION SYSTEMS

Error Bound EB. The small size preliminary sample is required in order to ini-tially estimate a mean positioning error xe for the positioning algorithm and theradiomap under examination. The confidence interval ci, indicates the probabilitythat the calculated mean positioning error provided by the distribution of sam-ples, describes the true radiomap positioning error and lies within the distributionspread. In other words, ci represents the desired confidence that the evaluatorwould like to have, and for this reason it is set by the evaluator himself at thebeginning of the process. Typical values of ci may be 67%, 95%, 99%, withoutlimiting the researcher in his work. The same applies to the Error Bound EB,which refers to the acceptable - in the evaluators’ opinion- error that may occur,relative to the calculated mean positioning error xe of the small size preliminarysample, always within the selected ci. These parameters are used as initial inputto the algorithm, assuming that the mean positioning error of the distribution ofsamples has a normal distribution (bell - curved). Based on the Central LimitTheorem, the aforementioned assumption is considered valid if the sample size issufficiently large. The population of all possible sample means can then be consid-ered approximately normally distributed, no matter what probability distributiondescribes the sampled population ([126]). The SSDA algorithm is presented inthe form of pseudo-code 3.2.1 and explained in more detail below.

Algorithm 3.2.1: SSDA(nsps, ci, EB,A)

procedure InitialPositioning(Positioning Algo, nsps)for i← 1 to nsps

do xe, sxe ← Positioning Algoreturn (xe, sxe)

maincomment: Step 1: Calculate df , t ci

2

df ← (nsps − 1)t ci

2← t-table

comment: Step 2: Estimate preliminary values of xe,sxe

for i← 1 to nspsdo InitialPositioning(WKNN, MMSE etc., nsps)

comment: Step 3: Calculate nSSS

nSSS ←(

t ci2sxe

EB

)2

comment: Step 4: Calculate GSmin

P ← (100 ∗ nSSS)

GSmin ←√A√

P+1

comment: Step 5: Select Random Samples

for i← 1 to nSSS

do Samplei ← Random x,y ∈ A, step GSmin

34

3.2. PROPOSED APPROACH

Initially, a small preliminary sample of size nsps that is randomly selected, isused for the initial rough estimation of the mean positioning error. The table ofthe standard normal curve cannot apply due to inefficient sample size, hence thet-distribution is implemented, which requires the estimation of degrees of freedomdf . From nsps, df can be calculated using the Equation 3.1:

df = nsps − 1 (3.1)

The t-distribution, given by Equation 3.2, in combination with the desiredci are used to estimate the t-value t ci

2for any small preliminary sample as per

equation 3.2.

f(tci) =Γ(df+1

2 )√dfπΓ(df

2 )

(1 +

t2

df

)− df+12

(3.2)

where Γ is the Gamma function.A more convenient way is to directly use the t-distribution table, available in

most probability and statistics books, based on which the tci value is given forseveral dfs. A sample of t-distribution table is depicted in figure 3.1.

Figure 3.1 – t-distribution table sample

35


The next step, involves an initial position estimation procedure based on thepreliminary sample data, in order to calculate the preliminary sample mean posi-tioning error xe, and the respective standard deviation sxe .

Assuming a normal distribution, as discussed earlier, the spread of the posi-tioning error, for the desired ci, is given by the following formula:

xe ± t ci2

(sxe√n

)(3.3)

where t ci2

(sxe√n

)is the maximum deviation from the mean value, hence the

error bound.Having set the desired EB as the basic evaluation criterion, the suggested

sample size (SSS) can be calculated by solving Equation 3.3 with respect to n:

nSSS =

(t ci

2sxe

EB

)2

(3.4)

In order to select an independent random sample of nSSS locations withinthe test environment, the population of the locations should be converted from aninfinite number (continuous) to discrete, in such a way that it fulfills the statisticalcriteria that the population P should be at least 100 times more than the samplesize ([127]). Hence,

P = 100nSSS (3.5)

Using Equation 3.5 and the dimensions of the test environment (width w xheight h), the minimum grid size GSmin can be determined:

GSmin =

√A√

P + 1(3.6)

where A is the Area of the test environment in m2.The proposed methodology concludes with the selection of nSSS simple ran-

dom locations within A, using the grid size GSmin as a step. The simple randomselection is performed by a method in which each collection of nSSS locations isequally likely to comprise the sample. In any other case, where the selection of thesample may involve a predefined pattern (sample of convenience), e.g. selection ofall nSSS samples from a specific room in the area of interest, the calculated meanpositioning error, may differ systematically from the actual error of the RTLS([126]).

SSDA implementation ensures that the selected sample will reflect the overallsystem performance and characteristics, within the desired confidence level and thepredefined acceptable mean localization error bound, for the utilized positioningalgorithm. In case of evaluating more than one positioning algorithms, then theSSDA should be applied to each one of them and the common suggested samplesize nSSS should be the largest extracted from all algorithms, in order to ensureconsistency and reliability of the performance results for all testing scenarios.

36

3.3. TEST ENVIRONMENT

3.3. Test Environment

The proposed algorithm was tested using two RSS fingerprint data sets, alsoknown as radiomaps. The first radiomap was generated through measurementsand the second through simulation. The SSDA was implemented initially on thefirst radiomap and the performance was tested for several sample sizes, up to40 samples. The second - simulated – radiomap – was generated and utilizedextensively for much larger sample sizes (up to 500), in order to further investigatethe reliability and convergence of the proposed algorithm. The test environmentwas an indoor area of 169m2. The wireless network consisted of 6 D-Link 802.11APs, as shown in Figure 3.2.

Figure 3.2 – Test Environment with APs and Radiomap as Generated in TruNETSimulator

The first radiomap was generated through measurements using an Android-based MS device (HTC Desire HD). Fingerprints were collected at 110 locationsat a step of 1m, at a height of 90cm. The device orientation was kept constantthroughout the measurement positions. Additionally, during the measurementprocedure no human or machine motion was allowed in the whole test environment,in order to minimize any dynamic fluctuations that may have affected the quality ofthe generated radiomap. At every measurement point, 30 district RSS samples (1sample/sec) were recorded and the mean RSS value for each location was extractedto formulate the radiomap. The RSS values in the radiomap ranged from -99 dBmto -34 dBm.

The second radiomap was created using TruNET wireless, fully deterministicsimulator, as shown in Figure 3.2. The same building structure and large furniturewere imported and configured using material constitutive parameters obtainedfrom literature [128], as shown in Table 3.1. A calibration procedure was carried

37


out as described in [129]. A high density receiver location layout was used in thesimulated scenario, by defining the receiver step distance to 10cm. This resultedin the generation of 16900 fingerprints in the dataset, allowing the selection andtesting of a wide range of sample sizes.

Table 3.1 – Material Constitutive Parameters of the Test Environment

Material El. Perm. (F/m) Loss Tang.

Concrete 3.9 0.23Wood 2 0.025Brick 5.5 0.03Metal 1 1000000Plasterboard 3 0.067Glass 4.5 0.007

3.4. Performance Evaluation

3.4.1. Implementing SSDA

A number of sample sizes was calculated by implementing SSDA in the testenvironment, for different acceptable Error Bounds, maintaining the desired con-fidence interval at 95% ( ci

2 = 0.025). Four small size preliminary samples ofnsps = 10 were randomly chosen within the test environment (df = 9, andt0.025 = 2.262). Additionally, a fifth sample of convenience of the same size(ncon = 10) was selected, in order to observe the probable differences and high-light the importance of selecting a simple random sample. This specific sampleconsisted of measurements taken explicitly from the southern area of the test en-vironment as shown in Figures 3.3 and 3.4 . The southern area was selected basedon the observation that the estimated local positioning error in this specific areawas consistently less than the average of the test environment, offering the op-portunity to highlight the risk of selecting such type of samples. All five smallsize preliminary samples were utilized, as defined by the SSDA algorithm, for aninitial positioning procedure, using a deterministic (WKNN) and a statistical po-sitioning algorithm (MMSE) on the generated through measurements radiomap.The results of xe and sxe , which are necessary for the estimation of nSSS , aredepicted in Table 3.2. A representative example of the positioning error vectorsfor both WKNN and MMSE referring to simple random sample sps3 is shown inFigures 3.6 and 3.7 .

The respective positioning error distribution is presented in Figure 3.8. Thesuggested sample size nSSS and the recommended minimum grid size GSmin percase, provided by SSDA, are depicted in Table 3.4.

3.4.2. Simple Random Sample vs Sample of Convenience

Based on the test results presented in Table 3.2, it is observed that both xe andsxe values of the sample of convenience differ noticeably from the respective values

38

3.4. PERFORMANCE EVALUATION

Figure 3.3 – Positioning Error Vectors-Sample of Convenience con5 with ncon = 10- WKNN

Figure 3.4 – Positioning Error Vectors-Sample of Convenience con5 with ncon = 10- MMSE

of all measured simple random samples. The mean positioning error recorded in thecase of choosing a sample of convenience is xecon = 1.49m, which is less than theminimum recorded positioning error of all other samples. Although the difference isrelatively small, when combined with the calculated standard deviation sxecon =0.76, may clearly mislead the evaluator to select smaller sample sizes since itdesignates an RTLS with a small position error distribution. This observation isreflected in Figure 3.5 vs Figure 3.8. In the aforementioned figures, one can observea positioning distribution error range between 0.5m−3.2m in the case of the sampleof convenience, while in the simple random sample case (sps3) the recorded range isapproximately between 0.5−4.0m with some isolated position estimations reaching

39


(a) WKNN (b) MMSE

Figure 3.5 – Positioning Error Distribution-Sample of Convenience con5 with nsps =10

Figure 3.6 – Positioning Error Vectors-Random Sample sps3 with nsps = 10 -WKNN

an error between 7.5m to 10m. From the above comparison it is understood that bynot selecting an appropriate sample type and size, representative of the population,the outcome of the calibration or evaluation of the RTLS would not be objective.

3.4.3. Testing on a Fingerprint Dataset Generated through Measure-ments

Initially, the performance of the SSDA was investigated for the radiomap gen-erated through measurements. WKNN and MMSE positioning algorithms wereimplemented for different number of sample sizes ranging from n = 5 to n = 40.

40


Figure 3.7 – Positioning Error Vectors-Random Sample sps3 with nsps = 10 - MMSE

(a) WKNN (b) MMSE

Figure 3.8 – Positioning Error Distribution-Random Sample sps3 with nsps = 10

The behaviour of the mean and Circular Error Probability (CEP50%, CEP67%and CEP95%) is shown in Figures 3.9 and 3.10 .

It is observed that, for small sample sizes (n = 5 to n = 30), the positioningerror fluctuates up to 0.5m for the mean value, 0.7m for the CEP95% and 1.15mfor the maximum error for this specific environment. The fluctuation tends tostabilize in a smaller range (< 0.4m) for sample sizes above 35 in this specificradiomap. The observations agree with the calculated nSSS which suggest a sizeof 48 samples in order to achieve an EB = ±0.4m. The recommended nSSS

ensures that the performance results can be reliably verified, if a different sample,of the same or larger size, is chosen. A typical simple random selection for theaforementioned scenario is presented in Figure 3.11.

41


Table 3.2 – Mean Error and Standard Deviation of Preliminary Samples- radiomapgenerated through Measurements

Algorithm Sample xe(m) sxeWKNN sps1 1.73 1.227

sps2 1.78 1.268sps3 1.82 1.264sps4 1.59 1.078con5 1.49 0.763

MMSE sps1 1.72 1.307sps2 1.68 1.371sps3 1.74 1.378sps4 1.54 1.178con5 1.50 0.798

Table 3.3 – Suggested Sample Size for different desired EB values using Sample ofConvenience– radiomap generated through Measurements

Algorithm EB (m) nSSS GSmin (m)

WKNN ±0.25 48 0.185±0.40 19 0.290±0.50 12 0.360

MMSE ±0.25 53 0.176±0.40 21 0.280±0.50 14 0.340

Table 3.4 – Suggested Sample Size for different desired EB values using SimpleRandom sps –radiomap generated through Measurements

Algorithm EB (m) nSSS GSmin (m)

WKNN ±0.25 120 0.120±0.40 48 0.185±0.50 30 0.235

MMSE ±0.25 180 0.10±0.40 70 0.15±0.50 45 0.19

3.4.4. Testing on a Fingerprint Dataset Generated through Simulation

In order to test SSDA with a larger range of sample sizes, a high resolutionradiomap was generated by TruNET wireless, a 3D ray tracing simulator. Thegrid size in this scenario was set at 10cm, allowing the utilization of SSDA forEB values less that 25cm as extracted from Table 3.4. A WKNN positioningalgorithm was then implemented for sample sizes ranging from 5 to 500. Themean, CEP50%, CEP67% and CEP95% values are presented in Figure 3.12.

42


Figure 3.9 – Radiomap generated through measurements: Effect of sample size onpositioning error for WKNN

Figure 3.10 – Radiomap generated through measurements: Effect of sample size onpositioning error for WKNN

It is noted that, depending on the sample size, the mean value may fluctuateup to 1.95m and 2.62m for CEP95%. It is also observed that the estimated meanstabilizes within a range of ±25cm for sample sizes near n ' 150. SSDA algorithmrecommends the selection of 170 random locations based on a preliminary sample

43


Figure 3.11 – Typical Simple Random Sample Selection of 48 locations for WKNNAlgorithm - Small Preliminary Sample No3

Figure 3.12 – radiomap generated by a high-resolution simulation: Effect of samplesize on positioning error using WKNN algorithm

of only 10 points, while ensuring the reliability of any presented performanceresults or training procedure. It is also proved that any sample size greater thanthe one suggested, will not affect the outcome, hence such an action will only addunnecessary load.

44

3.5. CONCLUSION

3.5. Conclusion

In this Chapter, SSDA algorithm is presented that allows the calculation of asafe margin sample size to be used during the training, calibration and performanceevaluation of fingerprint based localization systems. The proposed methodologysuggests the utilization of an initial preliminary sample selection, the definitionof an acceptable positioning error bound and a predetermined confidence inter-val. The suggested sample size is then extracted by converting the locations frominfinite to discrete and by setting a minimum grid size for the area of interest. Ad-ditionally, the importance of selecting a simple random sample is highlighted andcompared with a sample of convenience, demonstrating that in the latter case, theresults can vary systematically, leading to unreliable conclusions. Finally, SSDAwas tested with radiomaps that were generated through measurements and simu-lations. The outcome indicated that the estimated data sample size, objectivelycaptures the actual system’s positioning accuracy performance. The presentedwork contributes towards the standardization of RTLS evaluation procedures.

45

CHAPTER 4

Quality Evaluation of Fingerprint–based RTLS

4.1. Introduction

The positioning accuracy achieved by any RTLS including fingerprint-basedplatforms, is affected by various factors, such as the number of APs, the utilizedtechnology, the positioning algorithms, the device diversity, the complexity of theenvironment and its dynamic nature. No matter how the aforementioned factorsaffect the positioning accuracy, the quality of the fingerprint radiomap is of criticalimportance to the overall system performance. Even though the accuracy of local-ization can be estimated roughly during the online mode, from a design perspec-tive, it is very important to know the positioning performance of the fingerprintdatabase, in an offline mode, before the RTLS platform deployment. Estimatingthe performance at this stage, a service provider could take actions to improve thewireless network deployment. Improvement options include but are not limitedto: installation of additional APs, setup modifications, transmitting power or an-tenna orientation reconfiguration, offering of incentives to crowd-source membersto contribute additional data in the area of interest etc.. In this Chapter, the “Tol-erance Based - Normal Probability Distribution (TBNPD)” correlation algorithmis proposed. TBNPD evaluates the similarity of each pair of fingerprint entriesforming the radiomap and assesses their degree of correlation. In this way, theproposed algorithm, provides a reliable and generic metric for the characterizationof the quality of the fingerprint radiomap which can influence the performance ofthe RTLS.


The approach presented in [2], differs from already presented methods, sinceit is not directed towards bench-marking or utilization of predetermined datasets,nor towards evaluating the output of the positioning algorithms utilized duringthe online phase. It concentrates on the evaluation of the fingerprint input datawhich feeds the positioning algorithms. The method aims to directly and inde-pendently evaluate the quality of any fingerprint radiomap, using a single genericalgorithm. Since the fingerprints of a radiomap constitute the core data for allcalculations for location estimation [110], the radiomap quality can primarily de-termine the positioning performance of the RTLS. The proposed methodologyassesses and combines several factors, such as the impact of network setup, includ-ing changes in the geometry and the number of the APs, as well as the effect ofany RSS dynamic fluctuations that may occur. To achieve this task, the TBNPDcorrelation algorithm is introduced, which estimates the degree of correlation be-tween each pair of fingerprint entries and assesses their uniqueness. In other words,

47

CHAPTER 4. QUALITY EVALUATION OF FINGERPRINT–BASED RTLS

TBNPD correlation algorithm imprints each radiomap’s signature which indicateshow well the radiomap is expected to perform if used for positioning purposes.Combining the generation of multiple radiomaps through simulation, of the samearea of interest, with the implementation of the TBNPD algorithm for evaluationpurposes – that is comparing their signatures– allows us to predict the impactof each radiomap on the performance of the RTLS system, before deploymentand preliminary modify the network setup in order to achieve better positioningaccuracy.

4.2.1. Factors and Parameters Encapsulated

Before analyzing the implementation of the TBNPD correlation algorithm,the main parameters encapsulated in the algorithm are briefly introduced.

Firstly, the Minimum Distance Threshold (MDT ) is defined, as being the radiusaround each measurement or simulated point, where no correlation calculationsshall be performed by the TBNPD algorithm. MDT is a filtering parameter, aim-ing to adjust the minimum distance in the proximity of each location, where thealgorithm will not perform any correlation calculations. In this way, better com-putational performance is provided, especially when the density of the fingerprintsis high.

The second parameter introduced, is the RSS Tolerance (RSStol). It is defined asthe probable uncertainty range of the RSS values. In other words, RSS Tolerancerepresents the probable offset error in dB, within the area of interest. This typeof fluctuation may appear due to several dynamic factors, such as influence ofthe human body on the received signal of the MS, human and machinery motionin the area of interest, device diversity, any inaccuracies in the use of simulationetc. In this respect, the authors of [130], after testing 17 devices, reported anaverage standard fluctuation of RSS values ranging from 1dB to 12.3dB. Similarobservations were reported in [79], during the calibration procedure of a simulatedradiomap with respect to measurements. The above findings provide reasonableground that the RSS uncertainty can be significant and can affect the positioningaccuracy.

Finally, the TBNPD algorithm measures the impact on the quality of the ra-diomap, of any changes in the network setup. This is achieved by encapsulatingin the algorithm the RSS values and the standard deviation of these values, foreach and every AP of the network. The importance of the network setup in indoorposition estimation is highlighted in [131], where the authors proposed a mech-anism to measure the degree of the AP relevance to the localization process. Inour case, using the proposed TBNPD algorithm to compare different fingerprintradiomaps of the same area of interest, which are generated under different net-work setups or different RSS fluctuation scenarios, one can pinpoint the impact ofsuch changes on their overall quality as input data, and can estimate the impacton the platform’s positioning accuracy.

48


4.2.2. Radiomap Pre-processing

A pre-processing step was required for the collected radiomap data, in order toimplement the proposed algorithm and graphically generate the radiomaps signa-ture. For this purpose, a database was designed in a way to support several typesof radiomaps that are utilized by different RTLS. The structure of the database issuch that can accept a wide variety of obligatory and optional parameters. Oblig-atory parameters include the radiomap ID (specific site), AP’s ID and transmittedpower, the MS device type used to perform the measurements, the measurements’position and the measured RSS from each AP. Optional parameters include theantenna characteristics and orientation of both APs and MS devices, factors thatmay dynamically affect the radio signal etc. The data gathered, is then processedin order to form the measurement profile for each measurement sample. Finally,the mean RSS (RSS) is calculated, based on the number of measurement samplesthat belong to the same measurement profile.

The parameters consisting a measurement sample and a measurement profile areindicated in Table 4.1. Table 4.2 presents the parameters that characterize theTransmitters (APs) and MS devices. Parameters written in italics indicate thata relationship exists with other database tables, while the star sign (*) declaresobligatory parameters.

Table 4.1 – Measurement Sample and Measurement Profile Parameters

Measurement Sample Measurement Profile

*ID *ID

*Measurement Profile ID *Rx Absolute Position IDTemperature *Transmitter IDTime Stamp *Device ID*RSSI *Mean RSSI

Influence Factor ID Orientation x,y,z,w

4.2.3. TBNPD Correlation Algorithm

The TBNPD algorithm as explained previously, estimates the correlation levelbetween each pair of fingerprint entries forming the radiomap. The resulting Cor-relation Score (CSTBNPD) is calculated by modifying the normal probability dis-tribution formula in order to encapsulate the level of RSStol and the standard devi-ation of each contributing AP and dynamically adjust its spread accordingly. Thenormal probability distribution calculation, is based on an accepted assumptionthat each RSS measurement sample is independent from the rest measurements.The independence assumption offers the possibility to estimate one-dimensionaldensities rather than multivariate densities, taking into account that the indepen-dence is only assumed locally, hence at one location every time [65]. Taking intoconsideration the above, the Correlation Score (CSpairAB) of any pair of random

49


Table 4.2 – Transmitter and Device Parameters

Transmitter Device

*ID *ID

*Position x,y,z Model

*MAC Address *Sensitivity

*Power Antenna ID*Frequency Category IDBandwidth Brand ID

Antenna ID

Orientation x,y,z,wCategory IDTechnology IDBrand ID*isActive

fingerprint entries, A (xA, yA) and B (xB , yB), for any active AP (APi), withoutintroducing the RSStol parameter, is given by formula 4.1:

CSpairAB =1

σRSSAPi

√2πe− 1

2 (RSSB−RSSAσRSSAPi

)2

(4.1)

where random fingerprint entries A,B ∈ Arearadiomap, distance dAB > MDT

and σRSSAPiis the standard deviation of the RSS values observed from each APi

as calculated in equation 4.2:

σRSSAPi=

√√√√∑ni=1RSS

2

i −(∑n

i=1 RSSin

)2n− 1

(4.2)

where n is the number of fingerprint entries in the radiomap and RSSi ≥MSSensitivity.

The introduction of the RSStol parameter in formula 4.1, leads to the proposedTBNPD correlation algorithm. The respective SCTBNPD for a pair of randompoints A,B ∈ Arearadiomap, for any APi is given by equation 4.3:

CSTBNPDpairAB =1

σRSSAPi

√2πe− 1

2 (RSSdiffσRSSAPi

)2

(4.3)

where

RSSdiff =∣∣∣(RSSA

)APi−(RSSB

)APi

∣∣∣− 2RSStol (4.4)

In formula 4.4, RSSdiff ≥ 0. For RSSdiff < 0 the value is set to 0, since therange of the RSS values of the fingerprint entries A and B overlap, designating ahigh level of correlation.

The total correlation score for a pair of entries (CSTBNPDtotal), is the productof the correlation scores of all active APs:

50


CSTBNPDtotal =

m∏AP=1

(CSTBNPDpairAB

)(4.5)

Finally, the Correlation Scores are normalized, in order to be used as multiply-ing factors for the calculation of the weighted average distance error (Eweighted).Eweighted is a metric that quantifies the quality of the fingerprints in the radiomapunder evaluation, with respect to the final positioning accuracy. In more simplewords, Eweighted takes into consideration the correlation level between all pairs offingerprint entries, and converts this correlation into a probability of error thatthe RTLS may fall into. This is expressed by formula 4.6.

Eweighted =

[∑kpair=1 (dpair‖CSTBNPDtotal‖)

]∑k

pair=1 ‖CSTBNPDtotal‖(4.6)

where ‖CSTBNPDtotal‖ are the normalized correlation scores.

4.2.4. RET:Radiomap Evaluation Tool

To graphically present the final radiomap signature, a software tool dubbedRET: Radiomap Evaluation Tool was developed in Visual Studio and thefingerprint database explained in subsection 4.2.2 was interconnected in order toretrieve and process the radiomap data. RET, includes the following functionalityin the main menu as shown in figures 4.1 and 4.2.

Figure 4.1 – RET: Main Window Toolbar - Part 1

Figure 4.2 – RET: Main Window Toolbar - Part 2

(1) SQL Server: Drop-down list for selecting and connecting with the SQLServer, that the radiomaps were saved in.

51


(2) Site: Drop-down list for all available Sites. Upon selection of a validSQL Server a drop down list of the recorded sites on the database ispopulated. The user can select the site in order to access the relativeradiomap information.

(3) Device: Drop-down list of available devices. Each radiomap includesinformation on the device (s) utilized during the measurements or sim-ulation procedure. Selecting a device directly associates the fingerprintsand information retrieved, only by that specific device.

(4) Algorithm: Drop-down list of available evaluation algorithms. A set ofevaluation algorithms that were developed and tested, are provided aswell. The novel TBNPD algorithm is also included in this list.

(5) MDT: Input Text-box for MDT value. The Minimum Distance Thresh-old filtering parameter in m, is set. It is the radius around each mea-surement or simulated point, where no correlation calculations shall beperformed by the selected evaluation algorithm. Detailed description isprovided in 4.2.1.

(6) RSS Tolerance: Input Text-box for RSS Tolerance value. The RSSTolerance represents the fluctuations of the RSSI in dB, that may appear,due to the dynamic nature of the environment. It is one of the mostimportant parameters utilized in the TBNPD Algorithm.

(7) Mean RSSI: An action button that calculates the Mean RSSI at eachlocation within the area of interest. Calculation is based on the numberof measurements recorded into the database, for the selected device.

(8) Run Correlation: An action button that initiates the evaluation pro-cedure, based on the selected configuration and algorithm. It correlatesthe measurement profiles of all locations and extracts the evaluation pa-rameter’s metrics: Weighted Mean Distance Error and Mean CorrelationScore, including their standard deviation.

The results are also graphically presented in the following forms:

(1) Selected Location vs Others: A sliding bar is utilized to slide throughthe whole list of possible user locations (all recorded fingerprints that ex-ist in the dataset). Upon selection of any location, its color becomesblack to highlight it as a point of interest. Simultaneously, internal cal-culations take place utilizing look-up tables, in order to provide a visualperception, in green and red colour shades, of the correlation levels withall other locations. Deep green colors indicate high correlation whiledeep red colours low correlation. As it can be understood, green colourspresent higher probability for the localization algorithm to generate er-rors. A snapshot of this presentation is shown in 4.3.

(2) 2D and 3D Radiomap Signature: In a separate tab, the correlationsbetween all pairs is graphically presented in 2D and 3D graphs. Duringthis procedure, we record the frequency of the correlation scores of allpairs of fingerprints (CSTBNPD), with respect to the Euclidean Distancewhich determines the potential distance error (E). High density of highCSTBNPD in high E, designates poor quality radiomaps. High densityCSTBNPD in low E, defines a more reliable radiomap. Typical examplesare presented in figure 4.4.

52


Figure 4.3 – RET: Selected Location vs All

(a) 2D Signature (b) 3D Signature

Figure 4.4 – Graphical Presentation of radiomap Signatures

To better understand the above, we introduce an imaginary 3 x 3 EvaluationGrid -Figure 4.5- which is layered above a 2D or 3D Signature graph. In thisway, the distribution of the paired fingerprints can be divided in 9 major clus-ters depending on their CSTBNPD and the E values. The vertical axis of theEvaluation Grid represents the CSTBNPD value of each paired fingerprint and itshorizontal axis its E value. As the cloud of the results (the radiomap signature)is shifted towards the HIGH Correlation - HIGH Error direction, the quality ofthe radiomap becomes questionable, since large estimation errors are more likelyto occur. As the radiomap signature is shifted at the low and left grid squares-LOW Correlation - LOW Error-, the positioning results will be more accurateand precise.

A typical example is presented in Figure 4.6, where the radiomap quality is im-proved due to the increase of the AP’s number from 3 to 6. The increase ofthe AP’s lowers the correlation between the fingerprint pairs and improves thepositioning accuracy.

53


Figure 4.5 – 3x3 Evaluation Grid for the radiomap Signature

Figure 4.6 – Comparison of radiomap quality: signature shift to low left: LOWCorrelation - LOW Error

54

4.3. TEST ENVIRONMENT

4.3. Test Environment

The proposed algorithm was tested using the radiomaps generated both bymeasurements (Empirical) and simulation (Simulated), as explained earlier in 2.3.The wireless network utilized is composed by 6 D-Link 802.11 APs, deployed inan indoor laboratory environment of approximately 250m2. The study area andthe exact locations of the APs, are shown in Figure 4.7.

Figure 4.7 – Test Environment: Measurement points and APs.

4.3.1. Radiomap Generated Through Measurements

The radiomap in the first scenario was generated by measurements collectedat 110 equally-spaced (1m spacing) locations, at a constant height of 90cm, asdepicted in Figure 4.7. At every measurement point, 30 discreet RSS samples, witha rate of 1 sample/sec, were recorded using an Android-based MS device (HTCDesire HD). The device orientation during the measurement procedure is keptconstant, and it is the same for every location. The same Android device is usedduring testing in order to minimize any device diversity issues. The RSS valuesrecorded in the radiomap ranged from −99dBm to −34dBm. Null measurementsare filtered and replaced with a value of −150dBm, in order to be mathematicallymanageable by the evaluation and localization algorithms.

4.3.2. Radiomap Generated Through Simulation

The second radiomap was created using TruNET wireless, a 3D polarimetricdeterministic simulator. The building structure and furniture were imported andconfigured using material constitutive parameters obtained from literature [128]and presented in Table 4.3.

A calibration procedure was carried out as described in [129]. The same 110measurement locations used during the measurement campaign performed in thefirst scenario, were also defined as receiver locations in the simulated scenario. Anoticeable difference between the measurement campaign and the simulation, isthat in the latter, only one RSS value per Rx location is recorded, which is set as its

55


Table 4.3 – Material Constitutive Parameters of the Test Environment

Material Electrical Permittivity (F/m) Loss Tangent

Concrete 3.9 0.23Wood 2 0.025Brick 5.5 0.03Metal 1 1000000Plasterboard 3 0.067Glass 4.5 0.007

mean RSS value. On the contrary, in the measurement campaign, an intermediatestep of the averaging the samples RSS values and extracting the mean RSS value,exists.


The effectiveness of the proposed algorithm was assessed by observing the im-pact of predetermined factors, on the quality of radiomap signatures. The impactof the selected factors was extensively tested by the research community, and iswell supported by documentation. Their behaviour is considered as a benchmarkand it is used to verify the performance of TBNPD algorithm. The three selectedfactors are the following:

(1) Number of APs: Several studies, such as [113] and [132], demon-strated that the number and the placement of the APs, affect the po-sitioning accuracy of the RTLS. Hence, as the number of participatingAPs in the positioning procedure is increased, keeping the geometry ofthe existing APs the same, we expect the Eweighted value to decrease andvice versa.

(2) Geometry of wireless network: As the APs’ location become morealigned and their convex area becomes less, the localization accuracyworsens hence, the quality of the radiomap should be more poor [113].Another way to interpret the geometry factor, is to expect a better ra-diomap quality, when the center of gravity of the APs approximates thecenter of gravity of the floor plan.

(3) Dynamic Fluctuation of RSS: The dynamic fluctuation of the RSS isencapsulated in the algorithm as the RSStol parameter. It was noted inthe literature that the dynamic environment influences the RSS valuesrecorded both during the online and offline phases [115], [117], [63]. Itis expected, that as the RSStol parameter increases, the quality of theradiomap to degrade.

The outcome is also compared with the resulting positioning error, whenWKNN and MMSE localization algorithms are implemented.

56


4.4.1. Effect of the Number of APs

During the testing procedure, the TBNPD algorithm was evaluated versus anincreasing number of APs. The impact of 3 and 6 active APs was investigated bymaintaining the RSStol constant at ±2dB for all radiomaps. Results are summa-rized in Table 4.4. It was found that the quality of the fingerprint radiomaps, inboth experiments, improves as the number of APs increases, as demonstrated bythe respective Eweighted and σE values. A sample of the radiomap Signatures for3 APs and 6 APs is imprinted in figures 4.8 and 4.9.

Figure 4.8 – Radiomap Signature for 3 APs

Figure 4.9 – Radiomap Signature for 6 APs

In the case of 3 APs, the signature of which is illustrated in figure 4.8, itcan be observed that, a heavy density of fingerprint entries appears at very highcorrelation levels (> 0.85). The respective calculated Distance Error also indicatesa spread ranging from 3m to 6m. On the other hand, the radiomap signature of6 APs, recorded a much lower Distance Error -both as an absolute value as wellas a spread-. The range depicted starts from 1.5m to 2.5m at high correlationlevels as shown in 4.9. In this scenario, the large density of fingerprint entriescan be observed at a correlation level below 0.2. The aforementioned distributionindicates that high values of positioning error are unlikely or at least more rare tohappen. The same trend was recorded at the output level, by both localization

57


algorithms as indicated by the recorded positioning error presented in Figures 4.10and 4.11. It is also observed that both simulated and measured radiomaps performapproximately the same. The results are aligned with [113],[115] and [132].

Table 4.4 – Radiomap Quality Evaluation using TBNPD Algorithm: Number ofAPs: 3 APs vs. 6 APs at RSStol = ±2dB

radiomap APs Eweighted σE CSmean σCS

Empirical 3 4.60 2.46 0.425 0.3136 3.96 2.19 0.314 0.284

Simulated 3 4.61 2.47 0.439 0.3186 3.65 2.16 0.300 0.293

Figure 4.10 – Effect of APs number: MMSE Algorithm

Figure 4.11 – Effect of APs number: WKNN Algorithm

58


4.4.2. Effect of Wireless Network Geometry

The second part of the investigation examines the effect of a varying networkgeometry, by maintaining the RSStol constant at ±2dBm. Two different geome-tries utilizing 3 APs were assessed. The first was an approximate equilateral tri-angular deployment which covered the maximum possible area Areat1 of the floorplan. The aforesaid deployment was modified for the second scenario by changingonly 1 out of 3 previously selected APs. The new AP position was such, that anobtuse triangular deployment occurred. The new area covered Areat2 < Areat1 .Both areas are shown in Figure 4.12. The proposed TBNPD algorithm was imple-mented in both generated radiomaps and the results are presented in Table 4.5.We can observe an increase in the recorded Eweighted and the respective σE , as thegeometry changes from an equilateral to an obtuse shape. The aforementioned in-crease of the probable Distance Error can be also noticed visually on the respectiveradiomap signatures, as illustrated in Figures 4.13 and 4.14. The density of thefingerprint entries at high correlation levels (> 0.85) is significantly higher in thecase of obtuse deployment when compared with the respective correlation levelsof the equilateral geometry. For the same fingerprint entries of the obtuse deploy-ment, the probable Distance Error has a range from 3m up to 9m, which is a muchhigher spread than the 3m to 6m range observed in the first scenario. Moreover,it can be seen that, in the obtuse geometry, the wide spread of the calculatedDistance Error occurs across the whole range of correlation levels, while in theequilateral geometry a much more tide distribution is imprinted. The behaviourof the TBNPD correlation algorithm is also aligned with the findings recorded in[113], where similar scenarios were investigated.

Figure 4.12 – Equilateral and Obtuse Triangular Deployment

4.4.3. Effect of RSS Tolerance

The final part of the performance evaluation, was to investigate the behaviourof the fingerprint radiomap with respect to the RSS Tolerance (RSStol) rangechanges. The RSS fluctuations recorded in [115] and [117] are up to 12 dB,

59


Table 4.5 – Equilateral vs. Obtuse Triangular Deployment

radiomap Eweighted σE CSmean σCS

Empirical equilateral 4.60 2.46 0.425 0.313obtuse 5.06 2.87 0.445 0.330

Simulated equilateral 4.61 2.47 0.439 0.318obtuse 4.98 2.79 0.435 0.320

Figure 4.13 – radiomap Signature for 3 APs - Areat1

Figure 4.14 – radiomap Signature for 3 APs - Areat2

leading to an RSStol = ±6dB. During this experiment, the RSStol value wasgradually increased from ±2dB to ±6dB for all 6 APs, for both measured andsimulated radiomaps. The analytical results for the aforementioned scenarios arepresented in Table 4.6.

All results point out that as the RSStol increases from ±2dB to ±6dB,Eweighted and σE also increase, indicating that the quality of the radiomap de-grades. The gradual shift of the radiomap signatures towards higher probabledistance error is depicted in Figures 4.15 and Figures 4.16.

In simulated radiomaps, it is observed that the density of the fingerprint entrieswhich are experiencing high correlation is gradually increased as RSStol values areincreased. Moreover, the calculated distance error range gradually shifts from 1m-3m to 1m-6m for high correlation levels (> 0.85). A gradual shift towards higher

60


Table 4.6 – radiomap Quality Evaluation using TBNPD Algorithm: RSS ToleranceEffects - 6 APs Active

radiomap RSStol Eweighted σE CSmean σCS

Empirical ±2dB 3.95 2.19 0.314 0.284±4dB 4.59 2.42 0.453 0.330±6dB 5.16 2.60 0.594 0.338

Simulated ±2dB 3.65 2.16 0.300 0.293±4dB 4.41 2.44 0.416 0.335±6dB 5.07 2.65 0.547 0.343

(a) Empirical ±2dB (b) Simulated ±2dB

Figure 4.15 – Empirical and Simulated radiomap Signatures: Effect of RSStol ±2dB

(a) Empirical ±6dB (b) Simulated ±6dB

Figure 4.16 – Empirical and Simulated radiomap Signatures: Effect of RSStol ±6dB

probable distance error is observed even in low correlation levels. The aforesaidbehaviour clearly points toward worse performance of the radiomap, in scenarioswith severe RSS fluctuations. An identical behaviour is recorded in the radiomapsgenerated by measurements (Empirical).

In order to validate the impact of the RSS fluctuation in terms of distance er-ror at the output level (online phase), a random RSS fluctuation of up to ±6dBwas injected in the initial radiomaps (empirical and simulated) and the WKNNand MMSE were implemented again. The testing data was recorded through mea-surements during an online phase. The outcome depicted in figure 4.17 confirmsthat distorted radiomaps, due to dynamic changes in RSS level, influence thelocalization error significantly.

61


(a) WKNN (b) MMSE

Figure 4.17 – Effect of RSS Tolerance at the Output Level: WKNN and MMSE - 6APs Active

Although it was found that MMSE localization algorithm was influenced moreheavily than the WKNN when increasing the RSS tolerance, the up-trend of thepositioning error was still similar, indicating that the radiomap quality affects thefinal outcome in both cases.

4.5. Conclusion

RSS fingerprint-based positioning techniques have been in use due to their im-plementation simplicity and their ability to provide acceptable accuracy. However,the evaluation of the quality of the input data, and the selection of a well perform-ing radiomap has been a constant challenge for all RTLS. In this chapter, a novelalgorithm is presented, for evaluating fingerprint-based positioning systems. TheTBNPD correlation algorithm evaluates the input fingerprint data recorded duringthe off line phase, instead of evaluating the output of the positioning algorithmsduring the online phase, which occurs after the deployment of the RTLS. The pro-posed methodology provides the possibility to assess the quality of any radiomapwith respect to different geometrical deployments, number of active APs, and var-ious RSS fluctuation levels, during the design phase of an RTLS. The performanceof the proposed algorithm was extensively tested for radiomaps generated throughmeasurements and simulations. Results were compared and have been found tosupport the correct behaviour of the TBNPD algorithm. The proposed algorithmcan be utilized to evaluate different radiomaps, referring to the same building en-vironment and can be used to select the best performer to be utilized as the inputfor the RTLS platform, thus improving its positioning accuracy.

62

CHAPTER 5

Evaluation of already deployed PositioningSystems

5.1. Introduction

In Chapters 3 and 4, the research work was concentrated on selecting the op-timum sample size during the training/calibration procedure and on evaluatingthe quality of the fingerprint datasets before the deployment of the localizationplatform. The complexity of the actual performance of an RTLS, increases whenthe system is finally deployed in the area of interest, since additional factors re-lated to the dynamic nature of the environment affect the accuracy and precisionof the system. The impact of such factors varies depending on the actual en-vironment, the time of the day that the localization service was requested, thehuman-machine mobility within the area; all these factors behave dynamicallyand make it extremely difficult to accurately estimate the actual localization accu-racy of the system. In this chapter, an evaluation methodology is presented whichadopts a binomial approach and offers a robust and effective way to calculate theactual, real time, on the field, performance of the positioning platforms.


Different from all other approaches, the work in [3], proposes the adoption of abinomial distribution methodology for the investigation of the claimed accuracy ofdeployed positioning systems (e.g. claimed accuracy: ca ≤ 2m with a CumulativeDistribution Function (CDF)= 95%).

A binomial experiment consists of n identical trials and each trial results in suc-cess m (estimated positioning error epe ≤ ca) or failure f (epe ≥ ca). Theprobability of success on any trial is p and remains constant from trial to trial,hence the probability of failure q = 1 − p also remains constant. In any bino-mial experiment, the binomial random variable x will represent the numberof successes in n trials and p(x) the probability of obtaining x successes in n tri-als. The information obtained by the binomial experiment can be then comparedwith the specifications of the positioning system and justify or reject its claimedaccuracy. This concept does not require any input from the dataset utilized bythe RTLS solutions, hence it can be used for the evaluation of any type of po-sitioning system, independently of the algorithms and the technologies that areused. More specifically, the binomial distribution approach can be divided in twosubcategories:

63

CHAPTER 5. EVALUATION OF ALREADY DEPLOYED POSITIONINGSYSTEMS

5.2.1. Discrete Random Variable Methodology

The binomial experiment can be approached with the discrete random variablemethodology, since x may assume a finite number of possible values, or the possiblevalues can take the form of a countable list. In this case, p(x) is given by thefollowing formula:

p(x) =n!

x!(n− x)!px(1− p)n−x (5.1)

The binomial probability distribution of x is then the list of all possible valuesof x:

p(x) = p(0) + p(1) + p(2) + ...+ p(n) (5.2)

The Strength of an Evidence, in the case of a binomial experiment for theevaluation of a positioning system, is defined as the probability that at least mout of n number of trials passed the success threshold, accepting the claimed caand CDF :

P (x ≥ m) = P (x = m) + P (x = m+ 1) + ...+ P (x = n) (5.3)

5.2.2. Continuous Random Variable Methodology

The second method approximates the binomial distribution by using a normaldistribution. In order for this approximation to be valid, n must be large enoughand p far from zero and one. These conditions are met when the following ruleapplies [133]:

np ≥ 5 and n(1− p) ≥ 5

The normal curve (bell-shaped graph) is generated using the formula:

f(x) =1

σ√

2πe−

12 (x−µσ )2 (5.4)

where µ = np is the mean and σ =√npq is the standard deviation. The

standard deviation is the square root of the variance npq.It should be noted that, since a discrete approximation (the binomial) is beingapproximated by a continuous distribution (the normal), a continuity correctionmust be made, in order to assign an area under the curve, and for the calculationsto be valid. The continuity correction, is as simple as adding or subtracting 0.5 tothe discrete x-value, using table 5.1 to decide whether to add or subtract.

After implementing the continuity correction, the procedure is straightforward:(i) The z-score can be easily calculated for any x-value, since the mean value andthe standard deviation are known (µ = np is the mean and σ =

√npq). (ii) The

z-score defines a certain probability from the z-table that is available in typicalbooks of statistics, a sample of which is depicted in figure 5.1.

64

5.3. TEST ENVIRONMENT AND MEASUREMENT SETUP

Table 5.1 – Continuity Correction Table

IF USE

P (X = n) P (n0.5 < X < n+ 0.5)P (X > n) P (X > n+ 0.5)P (X ≤ n) P (X < n+ 0.5)P (X < n) P (X < n− 0.5)P (X ≥ n) P (X > n− 0.5)

Figure 5.1 – Sample of z-score table

5.3. Test Environment and Measurement Setup

To test the approach, an experimental RSS fingerprint-based positioning sys-tem was deployed within un area of approximately 250m2. Four different devices(Lenovo X100e, HTC Desire HD, Samsung Nexus S and Samsung Galaxy Tab)were used for the generation of the respective radiomaps, retrieving data from aWLAN network composed of 6 D-Link 802.11 APs. The RSS measurements werecollected at 110 equally-spaced (1m spacing) training locations, at a constantheight of 90cm, as depicted in Figure 5.2.

65


Figure 5.2 – Measurement points and APs in the Laboratory Premises

At each measurement point, 30 distinct RSS samples (1 sample/sec) wererecorded, ranging from -98dBm to -15dBm. Finally, two positioning algorithms(KNN for k = 5 and MMSE for σ = 6) were implemented. K-Nearest Neighbour(KNN), is an algorithm that estimates the unknown user location, as the average ofK locations, which have the shortest Euclidean distances between the observed RSSfingerprint and the respective mean value fingerprints in the radiomap [83]. TheMinimum Mean Square Error (MMSE), is a probabilistic approach that estimatesthe unknown user location, assuming that its value is equivalent to the weightedsum of the locations in the radiomap. The aforementioned weights are set equalto the conditional probability of each location in the radiomap, given the currentobserved (posterior) fingerprint [65]. The theoretically calculated accuracy resultsof the performed experiment, per device, and for each localization algorithm, arepresented in Table 5.2.

Table 5.2 – Calculated Accuracy per Device

(a) KNN with k = 5,CDF = 95%

Device ca(m)

Lenovo 5.2HTC 3.6

Nexus 4.0Galaxy 3.5

(b) MMSE with σ = 6,CDF = 95%

Device ca(m)

Lenovo 4.7HTC 3.2Nexus 3.4Galaxy 3.2

66


It can be observed, that, although the fingerprint locations are identical for alldevices, and the conditions collected were also the same (simultaneous collection ofall samples, common sample rate and sample size), the claimed mean positioningerror varies significantly. This observation provides the first evidence that the ra-diomaps may not have the same quality and could not be interchangeable betweenthe devices. Such a suspicion will be also investigated and validated during theimplementation of the proposed approach.


In order to assess the effectiveness of the proposed methodology using thebinomial distribution approach, two different scenarios were investigated. In thefirst case, no radiomap calibration was carried out when using all four devices ateach others radiomaps (device diversity scenario). In the second case a LinearTransformation (LT) procedure was performed for calibration purposes. In bothcases, the radiomap of each device, either calibrated or not, was tested with all fourdevices, with both KNN and MMSE positioning algorithms. The success decisionthreshold was set at the previously calculated positioning accuracy of each device(as per Table 5.2) and the probability of success on any trial was hypotheticallyaccepted to be equal to the CDF (95%) in order to properly define the binomialscenario. For each experiment, n = 40 independent trials were performed, on arandomly selected route (refer to Figure 5.2). The results were then analyzed withrespect to the strength of evidence, which represents the uniqueness of the randomsample that was used during the testing procedure.

5.4.1. No Radiomap Calibration

The final outcome of the binomial experiment in the case of “No RadiomapCalibration” is summarized in Tables 5.3 and 5.4. The analysis provides strongevidence (marked in gray colour in tables) that the Lenovo radiomap cannot op-erate within the set specifications, when any of the other three hand-held devicesis used as the Mobile Station (MS) for positioning purposes. The aforementionedobservation is also justified by the results in a vice versa way. These statements arevalid for both KNN and MMSE algorithms since during the trials a percentage of22.5−32.5% failure rate for KNN and 20.0−52.5% for MMSE was recorded, whichis well out of the 5% provided by the 95% CDF . Additionally, strong evidencesuggests that HTC and Galaxy devices do not have interchangeable radiomaps.The strength of evidence is very high, as indicated by the extremely low values0.018% to 3.7E− 6%, which reflect the possibility of error of the performed trials.

The aforementioned results are in-line with the observations in [79] , where thesame experimental positioning platform and devices are used. In the aforemen-tioned work, the authors analyzed directly the raw dataset of the backbone ra-diomaps, and performed online experiments. Their findings indicate that specificradiomaps could not perform adequately in a device diversity scenario, withoutthe implementation of Linear Transformation.

67


Table 5.3 – Outcome of Binomial Approach for the KNN Algorithm

Radiomap Device

Lenovo HTC Nexus Galaxy

Lenovo Failure (%) 00.0 35.0 22.5 05.0Evidence(%) 12.8 3.7E-7 0.01 27.7

HTC Failure (%) 32.5 00.0 05.0 12.5Evidence(%) 3.7E-6 12.8 27.7 03.4

Nexus Failure (%) 25.0 00.0 05.0 10.0Evidence(%) 0.018 12.8 27.7 09.0

Galaxy Failure (%) 30.0 22.5 05.0 02.5Evidence(%) 3.2E-5 0.01 27.7 27.0

Table 5.4 – Outcome of Binomial Approach for the MMSE Algorithm

Radiomap Device

Lenovo HTC Nexus Galaxy

Lenovo Failure (%) 02.5 45.0 20.0 05.0Evidence(%) 27.0 1.4E-11 0.06 27.7

HTC Failure (%) 52.5 02.5 10.0 20.0Evidence(%) 2.4E-15 27.0 09.0 0.06

Nexus Failure (%) 45.00 07.7 02.5 10.0Evidence(%) 1.4E-11 18.5 27.0 09.0

Galaxy Failure (%) 40.0 15.0 07.5 05.0Evidence(%) 2.8E-9 01.0 18.5 27.7

5.4.2. Radiomap Calibration using Linear Transformation (LT)

Radiomap calibration is considered to be a fundamental step in the process ofhandling the variations of measured RSS data from different devices, with an aimto achieve a cross-device consistent level of positioning accuracy. For the purposesof this work, we implemented the linear transformation which is basically a lineardata fitting of the mean RSS values of one device versus the other devices, ona specific fingerprint location, using the least squares method. The least squaresmethod provides the best approximate linear fitting of a general vectorial equation

A ~X = ~b, by multiplying both sides with the transpose matrix of A in order to findthe projections of all points on that line. The generic formula that provides theleast squares solution, in a vector form, is given in 5.5.

ATA ~X∗ = AT~b (5.5)

A sample result of this procedure presenting the radiomap before and afterlinear transformation, is shown in Figure 5.3 and Figure 5.4 respectively.

68


−100 −90 −80 −70 −60 −50 −40 −30 −20 −10

−100

−90

−80

−70

−60

−50

−40

−30

−20

−10

Nexus Mean RSS [dBm]

Gala

xyT

ab M

ean R

SS

[dB

m]

Fitting Model = a*x+ba = 0.60726b = −19.5515

fitted curve

Figure 5.3 – RSS Correlation Plot between Galaxy Tab and Nexus before LT

−100 −90 −80 −70 −60 −50 −40 −30 −20 −10

−100

−90

−80

−70

−60

−50

−40

−30

−20

−10

Nexus Mean RSS [dBm]

Gala

xyT

ab M

ean R

SS

[dB

m]

Fitting Model = a*x+ba = 1b = −4.1026e−007

fitted curve

Figure 5.4 – RSS Correlation Plot between Galaxy Tab and Nexus after LT

The outcome of the binomial experiment in the case of ”Radiomap Calibra-tion” is shown in Table 5.5. It is observed that after the implementation of lineartransformation, all the devices perform satisfactory on all calibrated radiomaps.The failure percentage lies in most cases within specifications for both position-ing algorithms, while in the cases of HTC and Galaxy using MMSE, the failurepercentage is marginally out of specifications but the Strength of Evidence is notenough to prove that observation.

69


Table 5.5 – Outcome of Binomial Approach for Calibrated Radiomaps

Radiomap All Devices

KNN MMSE

Lenovo Failure (%) 02.5 02.5Evidence(%) 27.0 27.0

HTC Failure (%) 02.5 07.5Evidence(%) 27.0 18.5

Nexus Failure (%) 02.5 02.5Evidence(%) 27.0 27.0

Galaxy Failure (%) 02.5 07.5Evidence(%) 27.0 18.5

5.5. Conclusion

In this Chapter a practical methodology for the evaluation of the declaredaccuracy of positioning platforms, is introduced. The proposed methodology isbased on the binomial distribution approach. It can be applied to already deployedpositioning systems, in order to investigate their performance under dynamic en-vironments and device diversity scenarios. The robustness of this method is basedon the fact that it can be used irrespective of the positioning philosophy, net-work technology and positioning algorithms, while at the same time requires noinput from the positioning database. The same approach can be implemented andtested for other types of positioning methodologies that utilize TOA, TOF, AOAetc. Additionally, since in this work the binomial experiment was used to evaluatethe effectiveness of linear transformation, it can be further expanded to test othercalibration techniques such as the use of wide kernel and the implementation ofHLF.

70

CHAPTER 6

Fusing Information for Improving IndoorPositioning Accuracy

6.1. Introduction

This Chapter contributes towards the second part of the thesis objectives,which refers to the improvement of indoor localization accuracy and general per-formance, through the fusion of available information and/or combination of het-erogeneous wireless technologies. All the proposed methods and algorithms arefocused on the fingerprinting technique, as this is the most widely used and mostpromising for indoor localization. In Section 6.2 a novel algorithm is proposedwhich combines RSS data from IEEE 802.15 (BLE) and IEEE 802.11 (Wi-Fi)technologies. In Section 6.3, the same concept presented in Section 6.2 is uti-lized, but modified accordingly for VLP technology instead of BLE. Finally, inSection 6.4, the directional antenna pattern characteristics are used for improvingthe quality of the fingerprint datasets and to provide higher accuracy in indoorfingerprint–based localization.

6.2. Combining IEEE 802.15 (BLE) and IEEE 802.11 (Wi-Fi)Technologies

6.2.1. Introduction to BLE

BLE was introduced as part of Bluetooth 4.0 specifications, allowing the de-vices to support both BLE and classic Bluetooth protocols simultaneously [104].The power efficiency of Bluetooth was especially created for IoT applications. Itallows devices to run for long periods on extremely low power sources, such ascoin-cell batteries or energy-harvesting devices.

BLE operates at 2.4GHz and uses Gaussian Frequency Shift Key (GFSK)modulation in 40 channels of 2MHz. Three of the channels, called advertisingchannels, are used to ensure connectivity with other nodes, while the remaining 37,are the data channels. BLE has a range of around 100m in an outdoor environment,a maximum data rate of 1Mbit/s and an application throughput up to 305 kbit/s[104]. Finally, it supports point-to-point and mesh networks.

iBeacon was developed by Apple in order to provide a higher level of locationawareness, by utilizing the BLE technology. iBeacon is a cross platform technol-ogy for both Android and iOS devices, able to support the BLE standard [134].Devices acting as beacons, generate iBeacon advertisements through which theyestablish a region around them. Android and iOS mobile devices receiving theadvertisements can determine the entrance and exiting border from each Beacon’s

71

CHAPTER 6. FUSING INFORMATION FOR IMPROVING INDOORPOSITIONING ACCURACY

region, can estimate the nearest beacon and approximate the distance between thetwo devices. The aforementioned advertisements contain three identifying fields,as described in [135]:

• UUID: Universally Unique Identifier is a 128 bits integer used as an ID forall beacons in an application

• Major: is a 16 bits integer, used to differentiate Beacons with the same UUID.• Minor: is a 16 bits integer used to further differentiate Beacons that have the

same UUID’s and Major values.

Due to their design philosophy, iBeacons are flexible in deployment and canbe used in mobile objects or to temporarily define a region and sub-regions. Thepossibilities introduced by BLE and iBeacon technologies in indoor positioning,are currently a topic of investigation from the research community. The relatedwork on this topic was presented in Chapter 2, Section 2.4.

6.2.2. Proposed Approach

Our approach in [4], is a two step procedure that foresees to take advantageof the broad usage of BLE devices and their characteristics, which include theshort range, and ability to provide distance information. As a first step, thealgorithm gathers general localization data from BLE devices deployed in an IoTenvironment. Then, in the second step, it uses the BLE information to optimize thedata retrieved from existing, popular and low cost, IEEE 802.11 RSS fingerprint-based indoor positioning systems, in order to improve the provided positioningaccuracy. Implementing this concept, a new enhanced KNN positioning algorithmwas developed (i − KNN), which is able to filter the initial Wi-Fi fingerprintdataset (radiomap), taking into consideration the proximity of the RSS fingerprintsto the BLE devices. By choosing to filter the initial dataset instead of simplycombining BLE and Wi-Fi fingerprints, i−KNN utilizes an optimised small sizesubset of possible user locations for the final position estimation.

The i −KNN algorithm uses the data transmitted from the i-Beacons con-cerning the estimated distance between a BLE device and the mobile user (mobiledevice or body sensor), as well as the information referring to the nearest i-Beaconof the whole BLE network. This data is used as an input to the filtering processesof the i−KNN in order to roughly estimate a probable area A which encloses theuser’s position. The aforementioned donut-shape area A, is formulated betweena minimum and a maximum radius (RBLEmin and RBLEmax) measured from acenter point, where the BLE device is located. The RBLEmin and RBLEmax valuesare calculated, taking into consideration a predefined tolerance (Tol) parameter,which accommodates any positioning error factors. As illustrated in Figure 6.1,area A is then used to screen the number of candidate fingerprints, down to asubset (S : {`1, . . . , `k}) extracted from the initial IEEE 802.11 fingerprint dataset(D : {`1, . . . , `j}Wi−Fi). The filtered fingerprint data subset S is finally used

as the optimized input, to typical indoor positioning algorithms (in our case theKNN). The proposed methodology serves two purposes: Firstly, achieving fast po-sitioning estimation due to the utilization of a fragment of the initial fingerprintdataset and secondly achieving improved positioning accuracy by constraining any

72

6.2. COMBINING IEEE 802.15 (BLE) AND IEEE 802.11 (WI-FI) TECHNOLOGIES

possible calculation errors within a very specific area A, where the user is actuallylocated. The latter is achieved due to the inherited short range of IEEE 802.15and its capability to identify the nearest i-beacon device to each mobile user. Forclarity, a self-explanatory pseudocode of the i−KNN filtering algorithm used tocalculate the subset S : {`1, . . . , `k}, is presented in Algorithm 6.2.1, and explainedin Subsection 6.2.2.1.

Figure 6.1 – Concept of i-KNN: BLE utilization for Radiomap Subset Generation

73


6.2.2.1. i − KNN Algorithm Explanation. The filtering algorithm, receivesas an input the Wi-Fi radiomap D : {`1, . . . , `j}Wi−Fi and BLE locations B :

{BLE1 . . . , BLEi}. A procedure named Fingerprints to BLE distance pre-calculates in advance distances between BLE devices and all fingerprint locationsfor easier estimation of the final output which is subset S. Calculated distancesare given at the form of a matrix L : {r`1,BLE1 . . . , r`j ,BLEi}. During the real-time positioning estimation, the algorithm initially scans for traceable BLE’s byrunning BLE Discovery (B) procedure, retrieves their parameters and sortsthem accordingly based on their Received Signal Strength. Upon retrieval of BLEinformation, i − KNN selects the nearest BLE device and utilizes the distanceinformation broadcast by the beacon (RangeBLEi) to additionally calculate theTolerance Tol level and designate the donut-shape area A, shown in Figure 6.1.Tol can be either a constant number as in this research work, or can be dynamicallycalculated as a percentage b% of the distance information: ±RangeBLEi ∗ b%. Inthe latter case, the closer the MS user is to the BLE, the smaller the Tol will be, andhence a smaller optimized data–set will be selected. This methodology is a moreoptimized option, since the BLE-MS user distance calculation is more reliable atsmaller distances. In either case, upon calculation of area A, the optimized data–set S : {`1, . . . , `k}, is extracted by utilizing the pre-calculated distances retrievedfrom L : {r`1,BLE1

. . . , r`j ,BLEi} matrix. Finally, instead of utilizing the heavyinitial radiomap D, the much smaller data–set S feeds the typical KNN methodsin order to estimate the location of the MS user.

6.2.2.2. i −KNN in Pseudocode Form. For clarity reasons the i −KNN ispresented below in the form of a pseudo-code:

Algorithm 6.2.1: BLE Filter(

D : {`1, . . . , `j}WiFi, B : {BLE1 . . . , BLEi})

procedure Fingerprints to BLE Distance(D,B)for i← 1 to nBfor j ← 1 to nDr`j ,BLEi

return (L : {r`1,BLE1 . . . , r`j ,BLEi})

procedure BLE Discovery(B)for i← 1 to nBLE

doif (BLEi ∈ B)

then

{Retrieve BLE ParametersSort BLEs based on Nearest

return (BLENearest:ID,RSSI,Range)

74


..maincomment: Calculate BLE - Fingerprints distances

output (Fingerprints to BLE Distance(D,B))comment: Calculate filtering criteria

output (BLE Discovery(B))while (BLENearest! = null)

do

for i← 1 to nBif (BLEi == BLENearest)

then

comment: Calculate Tolerance (Tol) in [m]

Tol← ±RangeBLEi ∗ b%comment: Calculate Subset of Radiomap D (S)

for j ← 1 to nLif ((RangeBLEi + Tol ≤ r`j ,BLEi)or (RangeBLEi − Tol ≥ r`j ,BLEi))

then `j ∈ Sreturn (S : {`1, . . . , `k})

6.2.3. Performance Evaluation

For the practical implementation and testing of the positioning algorithms, anAndroid fingerprint-based experimental localization platform (φmap) was utilized,providing configuration capabilities for several parameters related with KNN andi−KNN algorithms. The most important parameters include K value, number ofsamples recorded per point, time interval between each sample and Tolerance (Tol)value. φmap also provides the possibility to select between different radiomaps(generated by both actual measurements or simulations) and to upload a 2-Dblueprint of the study area for user friendly visualization of the user position. Thei−KNN algorithm was implemented in φmap, and the log files where retrievedand analysed for performance evaluation purposes. A snapshot of the applicationis illustrated in Figure 6.2.

The Wi-Fi Indoor Positioning system was fully deployed, and tests were per-formed at 12 randomly selected locations as described previously for the bench-mark (Wi-Fi only) case and the two hybrid (Wi-Fi and BLE) scenarios. Dataretrieval was performed with the MS user to be static, in order to retrieve a statis-tically adequate number of samples allowing a valid statistical analysis. However,both φmap platform and i −KNN algorithm can support real time moving MSusers. The experimental results concerning the positioning error of the platformutilizing only the IEEE 802.11 radiomap and implementing the typical KNN al-gorithm, is presented in Table 6.1. An average positioning error of e = 4.05m withstandard deviation σ = 2.13 is achieved for the specific environment under study.The findings of the benchmark case are aligned with the performance of othertypical Wi-Fi indoor localization systems, a comparison of which is illustrated in

75


Figure 6.2 – φ−map Localization Platform

Table 6.2. A general conclusion extracted from the aforementioned table, is thattypical RSS fingerprint-based positioning systems can provide an accuracy rangingfrom 3m to 7m, practically meaning a room level designation.

Table 6.1 – Positioning Error of Wi-Fi RSS Fingerprint-based Positioning System

Test Point eaverage(m) σ Radiomap size (%) Samples No

1 5.09 2.96 100 192 2.45 1.79 100 213 3.81 2.15 100 214 4.70 3.76 100 205 4.15 2.56 100 256 6.55 2.26 100 237 5.21 2.02 100 168 2.91 1.56 100 299 5.61 2.02 100 3710 3.50 1.58 100 5011 3.02 1.01 100 3012 2.58 1.72 100 45

Test results for the first hybrid scenario (existence of a single BLE device)are shown in Table 6.3. The positioning error e is improved to 3.07m with asmaller standard deviation of σ = 1.60, while the utilized fingerprint data–set sizeis reduced to an average of 67%. At 5 out of 12 test points the i−KNN algorithmutilized the total number of fingerprints as included in the initial radiomap, since

76


Table 6.2 – Positioning Error of Typical Indoor Wi-Fi Positioning Systems

System Accuracy/Error Methodology Complexity

TIX,[136] 5.4mlinear mapping ofRSS

light algorithm withAP modifications

EZ, [137] 2.0− 7.0m model based complex algorithm

SDM,[138] 3mlinear mapping ofRSS

light algorithm withsniffers

Zee, [139] 3m RSS fingerprintscombined with Ho-rus or EZ

LiFS, [140] 89% room level RSS fingerprintscomplex trainingphase

WILL, [141] 86% room level RSS fingerprintscomplex mapping ofvirtual floor

φmap,[4] 4.05mWi-Fi RSS finger-prints

light algorithm

φmap,[4] , 2.33mWi-Fi-BLE RSSfingerprints

light algorithm

AnyPlace,[142], 1.96m

Wi-Fi crowd-sourcing-basedfingerprint collec-tion

light algorithm

no BLE signal was identified in these specific locations, due to wall attenuationeffects. Although BLE signals could penetrate single plasterboard walls and doubleglass windows, they were heavily weakened when transmission occurred throughcement and brick walls. In those 5 test points, i−KNN algorithm did not haveany effect and φmap operated as a Wi-Fi only platform. During the secondhybrid scenario, the full deployment of the BLE devices ensured that locationinformation from at least one i-Beacon was retrieved in all 12 test locations. Testresults of this scenario are presented in Table 6.4. A radical improvement of thepositioning accuracy and optimization of the utilized data–set occurred. Morespecifically, the error was reduced to 2.33m for the same study area, indicatingan improvement of 42%. Standard deviation was also significantly reduced to1.22, depicting the much higher concentration of results around the mean valuesof positioning results. Additionally, the utilized data–set size was significantlyreduced, ranging between 12% and 37% depending on the test point. Takinginto consideration the comparison Table 6.2, it is observed that the proposedi −KNN algorithm over-performs the typical Wi-Fi localization platforms. Forincreased accuracy, the RTLS operators need to deploy a number of BLE systemsis such a geometry that can cover the maximum area of interest. Obviously, sucha deployment depends on the complexity of the indoor environment; open planspaces require less BLE devices than wall-separated areas.

Figures 6.3 and 6.4 provide a visual performance comparison for the first sce-nario, while Figures 6.5 and 6.6 refer to the outcomes of the second scenario. Whatis obvious from the graphs is that, the combination of Wi-Fi and BLE systems

77


Table 6.3 – Positioning Error of Combined BLE (single BLE) and Wi-Fi RSSFingerprint-based Positioning System


1 1.60 1.22 29 192 2.62 1.75 53 213 2.77 2.01 38 214 3.63 2.65 45 205 3.64 2.35 40 256 2.17 2.16 39 237 5.21 2.02 100 168 2.91 1.56 100 299 3.20 1.90 69 3710 3.50 1.58 100 5011 3.02 1.01 100 3012 2.58 1.72 100 45

Table 6.4 – Positioning Error of Combined BLE (all deployed BLEs) and Wi-Fi RSSFingerprint-based Positioning System


1 1.96 1.44 22 192 1.70 0.93 12 213 1.22 0.81 12 214 1.52 0.62 16 205 2.38 1.14 12 256 2.12 0.63 16 237 2.33 0.67 16 168 2.91 0.65 12 299 3.84 1.89 37 3710 3.17 0.79 22 5011 2.97 1.26 28 3012 1.87 0.55 19 45

in the proposed i − KNN algorithm constantly outperforms the simple KNNalgorithm, especially when the BLE deployment is such that can provide adequatesignal coverage in the study area. In such scenarios, accuracy is improved andpositioning results fluctuate much less, as indicated by the lower standard devia-tion. Data–set utilization is optimized to an average of 20% for typical scenariosand can be further improved if the set Tolerance Tol factor is further optimized toa minimum value. Overall, the findings provide hard evidence that the proposedi−KNN algorithm improves the computational and processing requirements, pro-vides faster and more accurate positioning and incurs lower power consumption.

78


Figure 6.3 – Positioning Error Comparison: Wi-Fi only vs single BLE and Wi-Fi

Figure 6.4 – Fingerprint Dataset Size Utilization: Wi-Fi only vs single BLE andWi-Fi

Figure 6.5 – Positioning Error Comparison: Wi-Fi only vs nearest BLE and Wi-Fi

79


Figure 6.6 – Fingerprint dataset Size Utilization: Wi-Fi only vs nearest BLE andWi-Fi

6.3. Combining Smart Lighting and Radio in Indoor Positioning

6.3.1. Introduction to VLP

A Visible Light Positioning (VLP) system is based on Visible Light Commu-nication (VLC). As the name suggests, VLC is using beams of visible light to sendthe information [143]. The main challenge of VLC systems - finding a source ofartificial light that can be easily modulated - was solved with the development ofhigh power light-emitting diodes (LEDs). Responding to the potential use of VLC,IEEE published IEEE 802.15.7 standard (Short-Range Optical Wireless Commu-nications), to meet the needs of the rapidly growing market [144]. Conceptually,a VLP system is graphically illustrated in Figure 6.7. It consists of a transmitter(TxSL) which is the Smart Light, the receiver (Rx) which is the mobile station(user smart phone, wearable device with a light sensor etc.) and the environment[145]. The transmitter emits at a power PSL which follows a certain radiationpattern (RE(φi)). The encoded light follows the well known optical geometry, i.etravels in a direct path (LOS), reflects, diffracts or scatters on different obstaclesand falls onto the receiver. In the case of indoor positioning, the channel can beconsidered flat, since no high data rates are expected in the case of VLP, mak-ing the channel response more simple and taking into consideration only the LOSpaths [146], [147]. The mobile station is equipped with either a photodiode oran image sensor that converts the optical signal back to an electrical signal. Thereceived power PR depends on the transmitted power PTx and radiation patternRE(φi), the AOA (θi), the Active Area (AR) related to the FOV and clear ceiling-user height hcl and the spectral response. If an optical concentrator (G(θi), oroptical filter (T (θi)) exists on the receiver, it is taken into consideration. Theaforementioned relations are expressed with the formulas 6.1 and 6.2:

PR =PTx

d2RE(φi)T (θi)G(θi)AR cos(θi) (6.1)

and

80

6.3. COMBINING SMART LIGHTING AND RADIO IN INDOOR POSITIONING

d = 4

√PTx

PR

h2

πAR (6.2)

Figure 6.7 – Angles φ and θ to compute channel impulse response and SL Radius

In order to achieve accurate localization in VLP systems, the mobile stationmust be able to receive and process signals from multiple smart lighting sources.This fundamental requirement is the origin of most of the challenges in this spe-cific research area. Several methods have been investigated depending on whetherthe mobile station uses an image sensor or a single photodiode. In the case of animage sensor, each smart light can be processed by a different part of the sensor,without any synchronization requirements, hence Spatial Division Multiple Access(SDMA) can be used [148], [149]. On the other hand, in the case of a single pho-todiode, either Time Division Multiple Access (TDMA) is implemented [150] orFrequency Division Multiple Access (FDMA) [145]. TDMA requires synchroniza-tion in order to manage the time slots for every smart light communication channeland both RSS and TOF parameters can be utilized for positioning. FDMA, doesnot require any backbone since each smart light is allocated a specific frequencycarrier, but only the RSS can be used. Code Division Multiple Access (CDMA) isanother option, where every smart light is assigned a unique sequence to encodeits information. However, several interference issues occur due to the fact thatnone of the used codes are fully orthogonal [151].

The aforementioned methods although they benefit from the inherent prop-erties of the VLC, at the same time they suffer from the physical propagationconstraints of the light and existing technological limitations. Obvious techno-logical challenges faced, are related to low sample rate process capabilities of thecommercial image sensors, high processing power requirements, increased powerconsumption, necessity for dense smart light grids and/or expensive dedicatedinfrastructure for synchronization purposes [146].

81



The proposed approach presented in [5], is similar to the approach describedin Section 6.2. The concept is to combine the existing popular and low costIEEE 802.11 RSS fingerprint-based indoor positioning systems, with a minimaldeployment of smart lighting in order to assist and optimize the positioning ac-curacy of the former. A new enhanced KNN positioning algorithm was developed(vlp−KNN), able to filter the data received by the smart lighting infrastructure,calculate the required subset of the initial fingerprint dataset, and finally estimatethe user position.

The minimal smart lighting deployment is environment specific, but always lessthan 1/3 of the total number of deployed smart lights required in VLPs. Thismeans that depending on the space segmentation, the number of required smartlights may vary. However, in all cases instead of deploying dense and expensivesmart lighting grids that are necessary for achieving continuous overlapping ofrays from several smart lights, a single light per room is required. In wide andopen plan environments, the required number depends on several factors such asthe smart lighting transmit power (PTx, the Receiver Field of View (FOV ) andsensitivity, the existence or not of concentrating/optical lens, installation heightetc. Recalling (Figure 6.7), typical receivers of 160 degrees FOV (FOV = 2 ∗ φ),when deployed at a clear ceiling-user height of 2.0m can cover an area of radius ofapproximately 11.5m (total coverage of 404 m2) assuming that the illuminationpower is adequate.

The concept of the proposed vlp−KNN algorithm is to utilize the data receivedfrom a local smart light (RSS, and/or TOF, AOA) in order to roughly estimatea probable area A enclosing the user’s position. The aforementioned donut-shapearea A, is formulated between a minimum and a maximum Radius (RSLmin andRSLmax) from the the Smart Light (SL) location, taking into consideration a prede-fined tolerance (Tol) to accommodate any positioning error factors. As illustratedin Figure 6.8, Area A is then used to narrow down the number of candidate fin-gerprints to a subset (S : {`1, . . . , `k}) extracted from the initial IEEE 802.11fingerprint dataset (D : {`1, . . . , `j}WiFi). The filtered fingerprint data subset S isfinally used as input, when implementing typical indoor positioning algorithms (inour case the KNN). The proposed methodology serves two purposes: Firstly, fastpositioning estimation due to the utilization of a fragment of the initial fingerprintdataset and secondly improved positioning accuracy by minimizing any possiblecalculation errors to a very specific area A, where the user is actually located.The latter achievement is a result of the VLP property to rely only on Line ofSight rays, avoiding multipath effects. For clarity, a self explanatory pseudocodeof the algorithm used to calculate the subset S : {`1, . . . , `k} is presented in Algo-rithm 6.3.1. Finally it is noted that all distances between SL’s and all fingerprintlocations are pre-calculated, for easier estimation of the subset S.

82


Figure 6.8 – Combined VLP and Wi-Fi Fingerprint Based Indoor Positioning

83


Algorithm 6.3.1: vlp-Filter(

D : {`1, . . . , `j}WiFi, SL : {SL1 . . . , SLi})

procedure Fingerprints to SL Distances(D,SL)

comment: Calculate distances between all ì and SLi

for i← 1 to nSL

for j ← 1 to n`r`j ,SLi

return (L : {rì,SL1. . . , r`j ,SLi})

procedure In Range SL Radius(SL)

for i← 1 to nSL

doif (SLi ∈ SL)

then

Retrieve In Range SL ParametersSLInRange, XSL, YSL, ZSL

AOASL, RSSISL

Calculate Estimated Smart Light Radius (RSL)RSL = ZSL−ZUser

tan (90−AOASL)

return (SLInRange, RSL)

main

comment: Calculate SL - Fingerprints distances

output (Fingerprints to SL Distances(D,B))comment: Calculate filtering criteria

output (In Range SL Radius(SL))while (SLi! = null)

do

for i← 1 to nSL

if (SLi == SLInRange)then

comment: Calculate Tolerance (Tol) in [m]

Tol← ±RSLi ∗ b%comment: Calculate Subset of Radiomap D (S)

for j ← 1 to nLif ((RSLi + Tol ≤ r`j ,SLi)or (RSLi − Tol ≥ r`j ,SLi))

then `j ∈ Sreturn (S : {`1, . . . , `k})

84



The Wi-Fi indoor positioning system was fully deployed, and tests were per-formed at 7 randomly selected locations, distributed across the whole study area,as shown with an X mark in Figure 6.8. At each test location both simple KNNand proposed vlp − KNN algorithms were implemented to estimate user posi-tion. The collected experimental results concern the same metrics, as in the caseof WiFi-BLE experiment: the positioning error and the percentage o the utilizedfingerprint dataset size. Table 6.5 imprints the Wi-Fi only case, where the wholefingerprint dataset is utilized without any screening. In Table 6.6 the combina-tion of Wi-Fi and VLP is depicted. The results are also graphically presented inFigures 6.9 and 6.10..

Table 6.5 – Positioning Error of Wi-Fi RSS Fingerprint-based Positioning System

Test Point Eaverage(m) σ RadiomapSize(%)

1 5.09 2.96 1002 2.45 1.79 1003 3.81 2.15 1004 4.70 3.76 1005 4.15 2.56 1006 6.55 2.26 1007 5.21 2.02 100

Table 6.6 – Positioning Error of Combined VLP and Wi-Fi RSS Fingerprint-basedPositioning System

Test Point Eaverage(m) σ RadiomapSize(%)

1 1.96 1.44 222 1.70 0.93 123 1.22 0.81 124 1.52 0.62 165 2.38 1.14 126 2.12 0.63 207 2.33 0.67 16

A significant improvement on the localization accuracy is observed, since, froma 4.7m average error, the proposed algorithm drops to 1.89m. The standard devia-tion σ also appears to be much smaller, pinpointing less fluctuations on the overallsystem performance. Finally, the utilized fingerprint dataset is reduced to approx-imately 20% of the initial dataset size. This finding means less computational andprocessing requirements, faster positioning and less power consumption.

85


Figure 6.9 – Positioning Error Comparison: WiFi only vs VLP-WiFi

Figure 6.10 – Fingerprint Dataset Size Utilization: Wi-Fi only vs VLP-WiFi

6.4. Directional Antennas for Improving Indoor Positioning

6.4.1. Introduction to the Utilization of Directional Antennas in Assist-ing Positioning

Directional antennas have been widely used in localization processes, mostlyinvolving Direction of Arrival (DoA) and Time of Arrival (ToA) positioning tech-niques. Authors of [152], presented a smart switched beam antenna, which enableddirection of arrival estimation, using six directional planar elements. Directional

86

6.4. DIRECTIONAL ANTENNAS FOR IMPROVING INDOOR POSITIONING

antennas adopting the DoA positioning technique were also investigated in [153].Localization was performed by combining a rotational directional antenna withthe calculation of the optimum signal sampling rate, in order to ensure the record-ing of the highest RSS value, which would lead to the location estimation of thetarget.

Additionally, researchers in [154] introduced directional antennas with smallside lobes in ToA based positioning systems, in an effort to decrease errors causedby multipath components. Finally, in [155], the problem of cooperative localiza-tion of primary users in cognitive networks is addressed. In the aforesaid case,localization is achieved from secondary users by obtaining DoA measurementsfrom the target primary user, by means of directional antennas, multi-antennareceivers, or virtual arrays.


In [6], directional antennas are introduced in RSS fingerprint-based positioningsystems, in an effort to decrease the positioning error. The aim is to position andorient a directional antenna in such a way, as to reduce the correlation level of theRSS fingerprints forming the radio–map, while at the same time not degrading thecoverage of the wireless network.

The formation of more distinct fingerprints, is expected to improve localizationestimations by reducing the positioning error.

In order to achieve the above, a radio–map was initially generated, with APsusing only omni–directional antennas. The positioning accuracy of the radio–map (mean error and CEP95%) was then estimated, in order to be utilized asa comparison benchmark. Furthermore, the RSS correlation level (differences ofRSS values) at each location was calculated, resulting in the identification of highcorrelation areas and probably areas with the highest positioning error. This steptakes also into consideration any possible RSS fluctuations that may occur due tothe dynamic nature of the environment by introducing an RSS fluctuation factor.Experiments performed in [79] and [130], suggest that typical RSS fluctuationsmay range up to 12dB. The APs serving the area with the highest RSS correlationlevel, are designated as probable candidates for installing a directional antenna.Each candidate is then tested for different antenna orientation setups, through thegeneration of a new radio–map. Finally, the respective positioning accuracy andthe RSS coverage, are compared with the benchmark radio–map and all others, inorder to decide which is the optimum one in terms of positioning accuracy. Thisrecursive procedure can be easily performed by utilizing 3D Ray Tracing (RT)methods.


The aforementioned approach was tested in the indoor environment shown inFigure 6.11.

The environment and the wireless network utilized, is the same as in previousexperiments, but with the introduction of directional antennas. It was initially

87


Figure 6.11 – Test Environment: Measurement points and APs.

composed of 6 D-Link 802.11 APs with omni–directional antennas. Typical ma-terial constitutive parameters were obtained from literature [128]. A calibrationprocedure was then carried out as described in [129]. The benchmark radio–map was generated in TruNET, a 3D ray tracing simulator, by calculating theRSS values at 110 equally-spaced (1m spacing) locations, at a height of 90cm. Amean positioning accuracy of 2.2m and a CEP95% of 4.4m was estimated for thisbenchmark radio–map.

Subsequently, the RSS correlation level at each location was calculated. AP#2 was selected for installing a directional antenna, based on its proximity to thearea with the highest positioning error. The directional antenna patterns utilizedare shown in Figure 6.12.

15 dB

10 dB

5 dB

0 dB

−5 dB

−10 dB

−15 dB

−20 dB

0°

30°

60°

90°

120°

150°

180°

210°

240°

270°

300°

330°

Azimuth

Elevation

Figure 6.12 – Directional Antenna: Vertical and Horizontal Cut Patterns.

Eight radio–maps were then generated in TruNET, for different directionalantenna orientations (ranging from 270◦ to 350◦ at a step of 10◦), taking into

88

6.4. DIRECTIONAL ANTENNAS FOR IMPROVING INDOOR POSITIONING

account the building structural constraints, for the same 110 locations, as in theinitial scenario.

The positioning procedure included a test sample of 100 randomly selectedsamples in the area of interest, and the utilization of two localization algorithms:WKNN from the deterministic category, and MMSE from the probabilistic cate-gory. The results are summarized in Tables 6.7 and 6.8 and presented graphicallyin Figure 6.13.

Table 6.7 – Positioning Error per Azimuth - WKNN

Error [m] 270◦ 280◦ 290◦ 300◦ 310◦ 320◦ 330◦ 340◦ 350◦

mean 2.27 1.74 1.61 1.56 1.82 1.74 1.91 1.77 1.73CEP50% 1.94 1.30 1.48 1.46 1.58 1.43 1.61 1.55 1.42CEP67% 2.50 1.87 1.89 1.80 1.97 1.93 1.98 2.09 1.95CEP95% 4.84 4.12 3.02 3.19 4.29 3.85 4.05 3.97 4.13

Table 6.8 – Positioning Error per Azimuth - MMSE

Error [m] 270◦ 280◦ 290◦ 300◦ 310◦ 320◦ 330◦ 340◦ 350◦

mean 2.07 1.52 1.62 1.54 1.72 1.75 1.86 1.97 1.80CEP50% 1.66 1.22 1.46 1.31 1.54 1.52 1.58 1.85 1.52CEP67% 2.32 1.68 1.91 1.63 2.00 2.01 2.08 2.26 2.06CEP95% 4.87 3.95 3.06 3.36 3.99 3.52 3.98 3.95 4.29

(a) Mean Positioning Error. (b) CEP95% Positioning Error

Figure 6.13 – Positioning Error per Azimuth

The outcome of the experiment, shows a 29% improvement on mean posi-tioning error, from 2.2m to 1.56m, and 31% on the CEP95%, 4.4m to 3.02m asillustrated in Figure 6.14.

89


Figure 6.14 – Error Improvement: Omni Vs. Omni & Directional Antennas

The aforesaid improvement is noted at 300◦ orientation of the directional an-tenna. The results are also supported by the findings concerning the correlationlevel improvement, of the respective radio–map when compared with the bench-mark radio–map, as depicted in Figure 6.15.

Figure 6.15 – Correlation Level Improvement: Omni Vs. Omni & Directional An-tennas

90

6.5. CONCLUSION

It can be also observed that, at 270◦, both the mean and CEP95% errorsare higher than the respective errors in the omni–directional scenario. The aboveobservations suggest that, when directional antennas are introduced in a wirelessnetwork, they should be carefully oriented, in order not to affect negatively theperformance of the positioning system.

Finally, the effects of introducing a directional antenna in the wireless net-work are investigated with respect to the impact on the RSS coverage. For thisreason, the maximum RSS value at each location for both scenarios was recorded.As it is shown in Figure 6.16, the range of the aforementioned values remains al-most unchanged, while the average RSS values slightly improves (−46.59dBm to−47.09dBm), suggesting that no degrade occurs to the overall radio coverage ofthis setup.

Figure 6.16 – Radio coverage of tested Radio–maps: Maximum RSS values perlocation.

6.5. Conclusion

In Chapter 6 three new approaches were introduced with an aim to improveindoor localization accuracy and performance. The Wi-Fi-BLE and WiFi-VLPcombination that produced the i−KNN and vlp−KNN localization algorithmsare based on the same concept: to take advantage of the benefits of each tech-nology and fuse the available information in order to enhance the performanceof fingerprint–based indoor localization. Both algorithms are novel and, to theknowledge of the author, no previous research work has been proposed before.Test results for both algorithms, show a very similar behaviour and improvementin performance. This can be explained theoretically, based on the fact that bothsupporting technologies (IEEE 802.15 and IEEE 802.17) are utilized to filter andsize down the same fingerprint dataset that is generated by the same IEEE 802.11wireless network. Hence, assuming that the effective range for both supportingtechnologies is considered to be the same (size of a typical room), the filtering

91


procedure will approximately provide the same fingerprint sub-dataset for the per-formance of localization algorithms. Computational performance was indicated bythe reduction of the utilized dataset size, to a range between 12% and 37% of theinitial radiomap size. A significant improvement in localization accuracy was alsorecorded and reached a 42%. The scalability of both i−KNN and vlp−KNN al-gorithm makes it ideal for advancing the performance of existing fingerprint basedRTLS systems.

Concerning the third proposed concept, that refers to the utilization of directionalantennas for improving the quality of fingerprint databases, it has an additionalvalue when it is utilized during the design phase - before the deployment of thelocalization platform. During this step, the designing team can generate a seriesof radiomaps, measure the quality of the data and performance (i.e through theimplementation of the TBNPD algorithm presented in 4) and design beforehandthe optimum wireless network setup and the antenna configuration. To com-pensate the necessity for generating various radio–maps under different antennaorientation conditions, the use of 3D ray tracing simulations should be adopted,as recommended in the respective research work.

92

CHAPTER 7

Conclusion and Future Work

7.1. Conclusion

People and asset location information is one of the most critical parametersfor the provision of a wide range of enhanced services to both military and civiliansectors. Real time tracking of friendly and enemy forces, asymmetric threadsand crisis management, first responders coordination in natural disasters, are justa few of the numerous applications, which require critical location informationof the involved assets and people in air, land and sea. Civilian services includealmost every sector of industry and every target segment: health, transportation,marketing, finance, commerce, energy, etc.

The rapid technological developments in the era of IoT and Body Sensor net-works, have elevated indoor localization to one of the most active and ongoing re-search topics, with the core challenge to provide accuracy improvement and reliableperformance. Additional aspects under investigation concern self-improvementand adaptation to dynamic changes, low energy consumption and minimizationof calculation complexity. All the aforementioned research areas blend in a mix-ture of heterogeneous wireless technologies and availability of radio and non-radioparameters.

The dominant wireless technologies utilized in indoor localization, have beenthe IEEE 802.11, and IEEE 802.15, due to their popularity and cost efficiency.When it comes to the dominant indoor localization technique, it it by far finger-printing, as supported by the research work presented in Chapter 2.

Within this very active research topic, one fundamental aspect that concernsall types of indoor positioning systems, is to objectively evaluate the perfor-mance of the approaches and methodologies implemented. The contribu-tion of this thesis towards a more standardized and reliable evaluation has beendisseminated through published work presented in Chapters 3, 4 and 5. Three newnovel methods have been proposed related to the evaluation of fingerprint-basedRTLS:

(1) Sample size determination and unbiased selection of training/ calibra-tion/ evaluation samples, [1].

(2) Early, from the design phase, quality evaluation of the backbone ra-diomaps, through the introduction of TBNPD algorithm [2].

(3) Evaluation of already deployed systems by customizing a binomial dis-tribution approach, [3].

93

CHAPTER 7. CONCLUSION AND FUTURE WORK

The second fold of this research, was related with the provision of improvedlocalization performance in indoor environments by combining informa-tion from different wireless technologies. In this research area, this thesiscontributed in developing and testing 2 new localization algorithms as per [4] and[5]; and one new methodology presented in [6]. The algorithms (i − KNN andvlp−KNN) rely on the concept of utilizing information from a supportive tech-nology for filtering and minimizing, down to a subset, the fingerprints utilized forpositioning, of a Wi-Fi RSS fingerprint dataset. Both algorithms provide promis-ing results as described in Chapter 6. The new methodology as described in thesame Chapter, involved the utilization of directional antennas in an effort to im-prove the quality of a radiomap, by reducing the correlation between fingerprintneighbouring pairs.

7.2. Future Work

Future work in such a multidimensional and significant topic is endless. How-ever, in this section, some of the most important and challenging areas spottedduring the journey of this thesis, will be highlighted, in order to drive future re-search contributions:

Setting a well accepted and standardized methodology for evaluatingthe performance of localization platforms is still an ongoing challenge. Sig-nificant contributions were provided in this direction, such as the SSDA and TB-NPD algorithms and the Binomial Distribution methodology. However furtherexploitation of the aforementioned methods can be performed, in order to estab-lish standardization and bench-marking in evaluation methodology of RTLS. Anobvious example refers to the topic of sample size selection procedure, which canbe investigated and expanded further from the typical simple random selection. Itis the authors personal belief that research work, novel approaches and techniquesshould be objectively evaluated, before presented by the research community. Asafe way is to define and follow a set of predetermined criteria, mathematicallyproved and valid, which will provide a strong evidence of the reliability and objec-tivity of the performance evaluation procedure and outcome.

Expanding and investigating the scope of factors influencing the perfor-mance of RTLS, seems to be a topic that will help the design of more reliablesystems. The introduction of new technologies, the development of the capabili-ties and functionality of the existing wireless networks and their dense coexistence,provides a probable explanation for the unexpected but rather frequent degradein the performance of the deployed RTLS. Towards this direction, the backbonedatabase of the RET tool, presented in Chapter 4 was designed in a modular wayand a series of currently dormant parameters are recorded for future use. Param-eters include, environmental and instrumental parameters such as temperature,humidity and magnetic fields, neighbouring or overlapping frequency bands andsignal strength, human influence factors etc. In this respect, future work foreseenincludes:

(1) Investigation of human body effect on 3D antenna pattern and localiza-tion accuracy

94

7.2. FUTURE WORK

(2) Effects of building material constitutive parameters on Localization ac-curacy using Ray tracing techniques

(3) Investigation of the environmental conditions variation in the perfor-mance of WSN and BSN networks when utilized in positioning

(4) Circular polarization effects in indoor positioning

Further development of enhanced and agile localization algorithmsthat will be self-adapting and evolving. The introduction of Artificial Intelligence,and reinforcement learning, in the process of deciding from the available bulk in-formation, which to be chosen, fused and utilized, based on a set of dynamicallyevolving weighted criteria, seems to promise a breakthrough in indoor locationprediction with an improved accuracy.

Finally, focusing on the two supportive technologies (BLE and VLP) presentedin this research work, future work may include:

(1) investigating an optimal BLE deployment and an in-depth analysis ofthe body effect and sensor orientation, in order to further optimize theindoor localization performance.

(2) Computational performance improvements as well as battery saving mayalso be investigated, utilizing different models of smart devices underintensive use of the i −KNN and vlp −KNN algorithms versus otherhybrid solutions.

(3) Extensive testing and investigation of the effects of smart light deploy-ment density and 3D geometry, in further optimizing the indoor local-ization performance.

95

Bibliography

[1] Loizos Kanaris, Akis Kokkinis, Giancarlo Fortino, Antonio Liotta, and Stavros Stavrou.Sample size determination algorithm for fingerprint-based indoor localization systems.

Computer Networks, 101:169–177, 2016.

[2] Loizos Kanaris, Akis Kokkinis, Antonio Liotta, and Stavros Stavrou. Quality of fingerprintradiomaps for positioning systems. In 2017 24th International Conference on Telecommu-

nications (ICT) (ICT 2017), Limassol, Cyprus, May 2017.

[3] Loizos Kanaris, Akis Kokkinis, Antonio Liotta, Marios Raspopoulos, and Stavros Stavrou.A binomial distribution approach for the evaluation of indoor positioning systems. In 20th

International Conference on Telecommunications (ICT 2013), Casablanca, Morocco, 2013.[4] Loizos Kanaris, Akis Kokkinis, Antonio Liotta, and Stavros Stavrou. Fusing bluetooth

beacon data with wi-fi radiomaps for improved indoor localization. Sensors, 17(4), 2017.

[5] Loizos Kanaris, Akis Kokkinis, Antonio Liotta, and Stavros Stavrou. Combining smartlighting and radio fingerprinting for improved indoor localization. In Networking, Sensing

and Control (ICNSC), 2017 IEEE 14th International Conference on, pages 447–452. IEEE,

2017.[6] L. Kanaris, A. Kokkinis, M. Raspopoulos, A. Liotta, and S. Stavrou. Improving rss

fingerprint-based localization using directional antennas. In The 8th European Conference

on Antennas and Propagation (EuCAP 2014), pages 1593–1597, April 2014.[7] Akis Kokkinis, Marios Raspopoulos, Loizos Kanaris, Antonio Liotta, and Stavros Stavrou.

Map-Aided fingerprint-based indoor positioning. In 2013 IEEE 24th International Sympo-

sium on Personal, Indoor and Mobile Radio Communications: Fundamentals and PHYTrack (PIMRC’13 - Fundamentals and PHY Track), pages 265–269, London, United King-

dom, September 2013.[8] A. Kokkinis, L. Kanaris, M. Raspopoulos, A. Liotta, and S. Stavrou. Optimizing route

prior knowledge for map-aided fingerprint-based positioning systems. In The 8th EuropeanConference on Antennas and Propagation (EuCAP 2014), pages 2141–2144, April 2014.

[9] Akis Kokkinis, Aristodemos Paphitis, Loizos Kanaris, Charalampos Sergiou, and Stavros

Stavrou. Demo: Physical and network layer interconnection module for realistic planningof iot sensor networks. In Proceedings of the 2018 International Conference on EmbeddedWireless Systems and Networks, EWSN ’18, pages 201–202, USA, 2018. Junction

Publishing.[10] S. Mumtaz, K. M. S. Huq, J. Rodriguez, S. Ghosh, E. E. Ugwuanyi, M. Iqbal, T. Dag-

iuklas, S. Stavrou, L. Kanaris, I. D. Politis, A. Lykourgiotis, T. Chrysikos, P. Nakou, and

P. Georgakopoulos. Self-organization towards reduced cost and energy per bit for futureemerging radio technologies - sonnet. In 2017 IEEE Globecom Workshops (GC Wkshps),pages 1–6, Dec 2017.

[11] Revision of the commission’s rules to ensure compatibility with enhanced 911 emergencycalling systems, 1994.

[12] Alan Bensky. Wireless Positioning Technologies and Applications. Artech House, Inc.,

Norwood, MA, USA, 2007.[13] Min Chen. Towards smart city: M2m communications with software agent intelligence.

Multimedia Tools and Applications, 67(1):167–178, 2013.

[14] D. Zhang, L.T. Yang, M. Chen, S. Zhao, M. Guo, and Y. Zhang. Real-time locating systemsusing active rfid for internet of things. Systems Journal, IEEE, PP(99):1–10, 2014.

97

BIBLIOGRAPHY

[15] Hao Ji, Lei Xie, Chuyu Wang, Yafeng Yin, and Sanglu Lu. Crowdsensing: A crowd-sourcing

based indoor navigation using rfid-based delay tolerant network. Journal of Network andComputer Applications, 52(0):79 – 89, 2015.

[16] D. Macagnano, G. Destino, and G. Abreu. Indoor positioning: A key enabling technology

for iot applications. In Internet of Things (WF-IoT), 2014 IEEE World Forum on, pages117–118, March 2014.

[17] H. Liu, H. Darabi, P. Banerjee, and J. Liu. Survey of wireless indoor positioning techniques

and systems. IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applicationsand Reviews), 37(6):1067–1080, Nov 2007.

[18] K. Al Nuaimi and H. Kamel. A survey of indoor positioning systems and algorithms. In

Proc. Int Innovations in Information Technology (IIT) Conf, pages 185–190, 2011.[19] Reza Zekavat and R. Michael Buehrer. IEEE Press Series on Digital and Mobile Commu-

nication. IEEE, 2012.[20] R. M. Alkan, H. Karaman, and M. Sahin. Gps, galileo and glonass satellite navigation

systems: Gps modernization. In Proceedings of 2nd International Conference on Recent

Advances in Space Technologies, 2005. RAST 2005., pages 390–394, June 2005.[21] M. Richharia and L.D. Westbrook. Satellite Systems for Personal Applications: Concepts

and Technology. Wireless Communications and Mobile Computing. Wiley, 2011.

[22] Jacek Januszewski. Visibility and geometry of global satellite navigation systems constel-lations. Artificial Satellites, 50:169–180, 12 2015.

[23] J. A. del Peral-Rosado, R. Raulefs, J. A. Lopez-Salcedo, and G. Seco-Granados. Survey of

cellular mobile radio localization methods: From 1g to 5g. IEEE Communications SurveysTutorials, 20(2):1124–1148, Secondquarter 2018.

[24] Q. D. Vo and P. De. A survey of fingerprint-based outdoor localization. IEEE Communi-

cations Surveys Tutorials, 18(1):491–506, Firstquarter 2016.[25] S. S. Cherian and A. N. Rudrapatna. Lte location technologies and delivery solutions. Bell

Labs Technical Journal, 18(2):175–194, Sept 2013.[26] A. Tahat, G. Kaddoum, S. Yousefi, S. Valaee, and F. Gagnon. A look at the recent wire-

less positioning techniques with a focus on algorithms for moving receivers. IEEE Access,

4:6652–6680, 2016.[27] M. A. Al-Ammar, S. Alhadhrami, A. Al-Salman, A. Alarifi, H. S. Al-Khalifa, A. Alnafessah,

and M. Alsaleh. Comparative survey of indoor positioning technologies, techniques, and

algorithms. In 2014 International Conference on Cyberworlds, pages 245–252, Oct 2014.[28] A.K.M.M. Hossain and Wee-Seng Soh. A comprehensive study of bluetooth signal pa-

rameters for localization. In Personal, Indoor and Mobile Radio Communications, 2007.PIMRC 2007. IEEE 18th International Symposium on, pages 1–5, Sept 2007.

[29] Zhenlong Song, Gangyi Jiang, and Chao Huang. A survey on indoor positioning tech-nologies. In Qihai Zhou, editor, Theoretical and Mathematical Foundations of ComputerScience, pages 198–206, Berlin, Heidelberg, 2011. Springer Berlin Heidelberg.

[30] Abdulrahman Alarifi, AbdulMalik Al-Salman, Mansour Alsaleh, Ahmad Alnafessah,Suheer Al-Hadhrami, Mai A. Al-Ammar, and Hend S. Al-Khalifa. Ultra wideband indoorpositioning technologies: Analysis and recent advances. Sensors, 16(5), 2016.

[31] B. Amer and A. Noureldin. Rss-based indoor positioning utilizing firefly algorithm in wire-less sensor networks. In 2016 11th International Conference on Computer EngineeringSystems (ICCES), pages 329–333, Dec 2016.

[32] Nasrullah Pirzada, M Yunus Nayan, Fazli Subhanc M Fadzil Hassan, and Muhammad Amir

Khan. Device-free localization technique for indoor detection and tracking of human body:A survey. Procedia - Social and Behavioral Sciences, 129(0):422 – 429, 2014. 2nd Interna-

tional Conference on Innovation, Management and Technology Research.[33] R. Harle. A survey of indoor inertial positioning systems for pedestrians. IEEE Commu-

nications Surveys Tutorials, 15(3):1281–1293, Third 2013.

[34] Gabriel Deak, Kevin Curran, and Joan Condell. A survey of active and passive indoorlocalisation systems. Computer Communications, 35(16):1939 – 1954, 2012.

[35] A.K.M. Mahtab Hossain and Wee-Seng Soh. A survey of calibration-free indoor positioning

systems. Computer Communications, 66(0):1 – 13, 2015.

98

BIBLIOGRAPHY

[36] Mikkel Kjærgaard. A taxonomy for radio location fingerprinting. In 3rd international con-

ference on Location-and context-awareness, pages 139–156. Springer-Verlag, 2007.[37] S. He and S.G. Chan. Wi-fi fingerprint-based indoor positioning: Recent advances and

comparisons. Communications Surveys Tutorials, IEEE, 18(1):466–490, Firstquarter 2016.

[38] I. Guvenc and C. Chong. A survey on toa based wireless localization and nlos mitigationtechniques. IEEE Communications Surveys Tutorials, 11(3):107–124, rd 2009.

[39] M. Mah, N. Gupta, and A. Agrawala. Pinpoint time difference of arrival for unsynchronized

802.11 wireless cards. In 2010 Proceedings IEEE INFOCOM, pages 1–5, March 2010.[40] M. Ciurana, F. Barcelo-Arroyo, and F. Izquierdo. A ranging method with ieee 802.11 data

frames for indoor localization. In 2007 IEEE Wireless Communications and Networking

Conference, pages 2092–2096, March 2007.[41] S. A. Zekavat, A. Kolbus, X. Yang, Z. Wang, J. Pourrostam, and M. Pourkhaatoun. A

novel implementation of doa estimation for node localization on software defined radios:Achieving high performance with low complexity. In 2007 IEEE International Conference

on Signal Processing and Communications, pages 983–986, Nov 2007.

[42] R. Schmidt. Multiple emitter location and signal parameter estimation. IEEE Transactionson Antennas and Propagation, 34(3):276–280, March 1986.

[43] A. Barabell. Improving the resolution performance of eigenstructure-based direction-finding

algorithms. In ICASSP ’83. IEEE International Conference on Acoustics, Speech, andSignal Processing, volume 8, pages 336–339, April 1983.

[44] R. Roy and T. Kailath. Esprit-estimation of signal parameters via rotational invariance

techniques. IEEE Transactions on Acoustics, Speech, and Signal Processing, 37(7):984–995, July 1989.

[45] Janis Werner, Jun Wang, Aki Hakkarainen, Nikhil Gulati, Damiano Patron, Doug Pfeil,

Kapil Dandekar, Danijela Cabric, and Mikko Valkama. Sectorized antenna-based doa es-timation and localization: Advanced algorithms and measurements. IEEE Journal on Se-

lected Areas in Communications, pages 1–1, 11 2015.[46] A. Ishimaru. Electromagnetic Wave Propagation, Radiation, and Scattering. Prentice Hall,

1991.

[47] A. Aragon-Zavala. ANTENNAS AND PROPAGATION FOR WIRELESS COMMUNI-CATION SYSTEMS, 2ND ED. Wiley India Pvt. Limited, 2008.

[48] M.F. Catedra and J. Perez. Cell Planning for Wireless Communications. Artech House

mobile communications library. Artech House, 1999.[49] Y. Okumura, E. Ohmori, T. Kawano, and K. Fukuda. Field strength and its variabil-

ity in VHF and UHF land-mobile radio service. Review of the Electrical CommunicationLaboratory, 16(9-10):825–873, 1968.

[50] M. Hata. Empirical formula for propagation loss in land mobile radio services. IEEE Trans-actions on Vehicular Technology, 29(3):317–325, Aug 1980.

[51] European Cooperation in the Field of Scientific and Technical Research. Digital mobileradio: COST 231 view on the evolution towards 3rd generation systems. 1996.

[52] J. D. Parsons and M. F. Ibrahim. Signal strength prediction in built-up areas. part 2:Signal variability. IEE Proceedings F - Communications, Radar and Signal Processing,

130(5):385–391, August 1983.[53] K. Allsebrook and J. D. Parsons. Mobile radio propagation in british cities at frequencies

in the vhf and uhf bands. IEEE Transactions on Vehicular Technology, 26(4):313–323, Nov

1977.

[54] J. Walfisch and H. L. Bertoni. A theoretical model of uhf propagation in urban environ-ments. IEEE Transactions on Antennas and Propagation, 36(12):1788–1796, Dec 1988.

[55] S. Phaiboon. An empirically based path loss model for indoor wireless channels in lab-oratory building. In TENCON ’02. Proceedings. 2002 IEEE Region 10 Conference onComputers, Communications, Control and Power Engineering, volume 2, pages 1020 –

1023 vol.2, oct. 2002.[56] H. Hashemi. The indoor radio propagation channel. Proceedings of the IEEE, 81(7):943–

968, July 1993.

[57] D. Molkdar. Review on radio propagation into and within buildings. IEE Proceedings H -Microwaves, Antennas and Propagation, 138(1):61–73, Feb 1991.

99

BIBLIOGRAPHY

[58] A. Hills, J. Schlegel, and B. Jenkins. Estimating signal strengths in the design of an indoor

wireless network. Wireless Communications, IEEE Transactions on, 3(1):17 – 19, jan.2004.

[59] Manuel F. Catedra and Jesus Perez. Cell Planning for Wireless Communications. Artech

House, Inc., Norwood, MA, USA, 1st edition, 1999.[60] M. A. Youssef, A. Agrawala, and A. Udaya Shankar. Wlan location determination via

clustering and probability distributions. In Proceedings of the First IEEE International

Conference on Pervasive Computing and Communications, 2003. (PerCom 2003)., pages143–150, March 2003.

[61] MetaGeek LLC. inSSIDer Plus, https://www.metageek.com/products/inssider/, 2019.

[62] Fractal Networx Limited. TruNET wireless, www.fractalnetworx.com, 2017.[63] W. Wang, W. Wang, and K. Sun. Hybrid indoor localization method based on signal

subspace fingerprint database. In 2017 IEEE 17th International Conference on Commu-nication Technology (ICCT), pages 1132–1135, Oct 2017.

[64] Binghao Li, James Salter, Andrew G Dempster, Chris Rizos, et al. Indoor positioning

techniques based on wireless lan. In LAN, First IEEE International Conference on WirelessBroadband and Ultra Wideband Communications, pages 13–16, 2006.

[65] Teemu Roos, Petri Myllymaki, Henry Tirri, Pauli Misikangas, and Juha Sievanen. A prob-

abilistic approach to wlan user location estimation. International Journal of Wireless In-formation Networks, 9:155–164, 2002. 10.1023/A:1016003126882.

[66] Moustafa Youssef and Ashok Agrawala. The horus wlan location determination system.

In Proceedings of the 3rd international conference on Mobile systems, applications, andservices, MobiSys ’05, pages 205–218, New York, NY, USA, 2005. ACM.

[67] W. Kang and Y. Han. Smartpdr: Smartphone-based pedestrian dead reckoning for indoor

localization. IEEE Sensors Journal, 15(5):2906–2916, May 2015.[68] Y. Cheng, Q. Meng, Y. Liu, M. Zeng, L. Xue, and S. Ma. Fusing sound and dead reckon-

ing for multi-robot cooperative localization. In 2016 12th World Congress on IntelligentControl and Automation (WCICA), pages 1474–1478, June 2016.

[69] Seongwoo Jang, Kyungjae Ahn, Jongseong Lee, and Yeonsik Kang. A study on integration

of particle filter and dead reckoning for efficient localization of automated guided vehicles.In 2015 IEEE International Symposium on Robotics and Intelligent Sensors (IRIS), pages

81–86, Oct 2015.

[70] P. Bahl and V.N. Padmanabhan. RADAR: an in-building RF-based user location and track-ing system. In IEEE International Conference on Computer Communications INFOCOM,

volume 2, pages 775–784, 2000.[71] P. Prasithsangaree, P. Krishnamurthy, and P.K. Chrysanthis. On indoor position location

with wireless lans. In Personal, Indoor and Mobile Radio Communications, 2002. The13th IEEE International Symposium on, volume 2, pages 720–724 vol.2, Sept.

[72] TP Deasy and WG Scanlon. Simulation or measurement: The effect of radio map cre-ation on indoor WLAN-Based localisation accuracy. Wireless Personal Communications,42(4):563–573, 2007.

[73] Widyawan, M. Klepal, and D. Pesch. Influence of predicted and measured fingerprint on

the accuracy of rssi-based indoor location systems. In Proc. 4th Workshop Positioning,Navigation and Communication WPNC ’07, pages 145–151, 2007.

[74] Yiming Ji, Saad Biaz, Shaoen Wu, and Bing Qi. Impact of building environment on the per-

formance of dynamic indoor localization. In Proc. IEEE Annual Wireless and Microwave

Technology Conf. WAMICON ’06, pages 1–5, 2006.[75] M. Hassan-Ali and K. Pahlavan. A new statistical model for site-specific indoor radio prop-

agation prediction based on geometric optics and geometric probability. Wireless Commu-nications, IEEE Transactions on, 1(1):112 –124, jan 2002.

[76] Yiming Ji, Saad Biaz, Santosh Pandey, and Prathima Agrawal. ARIADNE: a dynamic in-

door signal map construction and localization system. In 4th ACM international conferenceon Mobile systems, applications and services, pages 151–164, 2006.

[77] A. Hatami and K. Pahlavan. Comparative statistical analysis of indoor positioning using

empirical data and indoor radio channel models. In 3rd IEEE Consumer Communicationsand Networking Conference (CCNC), volume 2, pages 1018–1022. IEEE, 2006.

100

BIBLIOGRAPHY

[78] K. El-Kafrawy, M. Youssef, A. El-Keyi, and A. Naguib. Propagation modeling for accurate

indoor wlan rss-based localization. In Proc. IEEE 72nd Vehicular Technology Conf. Fall(VTC 2010-Fall), pages 1–5, 2010.

[79] M. Raspopoulos, C. Laoudias, L. Kanaris, A. Kokkinis, C.G. Panayiotou, and S. Stavrou.

3d ray tracing for device-independent fingerprint-based positioning in wlans. In PositioningNavigation and Communication (WPNC), 2012 9th Workshop on, pages 109–113, March.

[80] Marios Raspopoulos, Christos Laoudias, Loizos Kanaris, Akis Kokkinis, Christos

Panayiotou, and Stavros Stavrou. Cross device fingerprint-based positioning using 3D raytracing. In Mobile Computing Symposium (IWCMC2012-Mobile Computing), Limassol,

Cyprus, August 2012.

[81] V. Honkavirta, T. Perala, S. Ali-Loytty, and R. Piche. A comparative survey of WLANlocation fingerprinting methods. In 6th Workshop on Positioning, Navigation and Com-

munication (WPNC), pages 243–251, 2009.[82] Wilson Yeung, JunYang Zhou, and Joseph Ng. Enhanced fingerprint-based location esti-

mation system in wireless LAN environment. Emerging Directions in Embedded and Ubiq-

uitous Computing, pages 273–284, 2007.[83] P Bahl, VN Padmanabhan, and A Balachandran. Enhancements to the RADAR user

location and tracking system. Microsoft Research, Tech. Rep. MSR-TR-00-12, February

2000.[84] S. Saha, K. Chaudhuri, D. Sanghi, and P. Bhagwat. Location determination of a mobile

device using ieee 802.11b access point signals. In 2003 IEEE Wireless Communications

and Networking, 2003. WCNC 2003., volume 3, pages 1987–1992 vol.3, March 2003.[85] H. Laitinen, J. Lahteenmaki, and T. Nordstrom. Database correlation method for GSM

location. In 53rd IEEE Vehicular Technology Conference (VTC2001-Spring), volume 4,

pages 2504–2508, 2001.[86] P. Kemppi and S. Nousiainen. Database correlation method for multi-system positioning.

In 63rd IEEE Vehicular Technology Conference (VTC2006-Spring), volume 2, pages 866–870, 2006.

[87] M.A. Youssef, A. Agrawala, and A. Udaya Shankar. WLAN location determination via

clustering and probability distributions. In IEEE International Conference on PervasiveComputing and Communications (PerCom), pages 143–150, 2003.

[88] M. Youssef and A. Agrawala. Handling samples correlation in the horus system. In INFO-

COM 2004. Twenty-third AnnualJoint Conference of the IEEE Computer and Communi-cations Societies, volume 2, pages 1023 – 1031 vol.2, march 2004.

[89] Moustafa Youssef and Ashok Agrawala. The Horus WLAN location determination system.In 3rd ACM International Conference on Mobile systems, applications, and services, pages205–218, 2005.

[90] M. Bshara, U. Orguner, F. Gustafsson, and L. Van Biesen. Fingerprinting localization inwireless networks based on received-signal-strength measurements: A case study on wimaxnetworks. IEEE Transactions on Vehicular Technology, 59(1):283–294, Jan 2010.

[91] S. Bai and T. Wu. Analysis of k-means algorithm on fingerprint based indoor localizationsystem. In 2013 5th IEEE International Symposium on Microwave, Antenna, Propagation

and EMC Technologies for Wireless Communications, pages 44–48, Oct 2013.[92] T. Chuenurajit, S. Phimmasean, and P. Cherntanomwong. Robustness of 3d indoor local-

ization based on fingerprint technique in wireless sensor networks. In 2013 10th Interna-

tional Conference on Electrical Engineering/Electronics, Computer, Telecommunications

and Information Technology, pages 1–6, May 2013.[93] Q. Chen and B. Wang. Finccm: Fingerprint crowdsourcing, clustering and matching for

indoor subarea localization. IEEE Wireless Communications Letters, 4(6):677–680, Dec2015.

[94] J. Zhang, G. Han, N. Sun, and L. Shu. Path-loss-based fingerprint localization approach

for location-based services in indoor environments. IEEE Access, 5:13756–13769, 2017.[95] Z. E. Khatab, A. Hajihoseini, and S. A. Ghorashi. A fingerprint method for indoor local-

ization using autoencoder based deep extreme learning machine. IEEE Sensors Letters,

2(1):1–4, March 2018.

101

BIBLIOGRAPHY

[96] Y. Kim and S. Kim. Design of aging-resistant wi-fi fingerprint-based localization system

with continuous active learning. In 2018 20th International Conference on Advanced Com-munication Technology (ICACT), pages 1054–1059, Feb 2018.

[97] J. Zuo, S. Liu, H. Xia, and Y. Qiao. Multi-phase fingerprint map based on interpolation for

indoor localization using ibeacons. IEEE Sensors Journal, 18(8):3351–3359, April 2018.[98] C. U. Park, H. Shin, and Y. Choi. A parallel artificial neural network learning scheme based

on radio wave fingerprint for indoor localization. In 2018 Tenth International Conference

on Ubiquitous and Future Networks (ICUFN), pages 794–797, July 2018.[99] A. Zhang, L. Guo, Q. Wu, and Q. Zeng. Fingerprint database optimization method for

indoor localization based on neighbor mean filter. In 2018 7th International Conference

on Agro-geoinformatics (Agro-geoinformatics), pages 1–6, Aug 2018.[100] D. Macagnano, G. Destino, and G. Abreu. Localization with heterogeneous information. In

Internet of Things (WF-IoT), 2014 IEEE World Forum on, pages 124–129, March 2014.[101] Jose Rabadan, Victor Guerra, Rafael Rodrguez, Julio Rufo, Martin Luna-Rivera, and

Rafael Perez-Jimenez. Hybrid visible light and ultrasound-based sensor for distance esti-

mation. Sensors, 17(2), 2017.[102] G. Fortino, R. Giannantonio, R. Gravina, P. Kuryloski, and R. Jafari. Enabling effective

programming and flexible management of efficient body sensor network applications. IEEE

Transactions on Human-Machine Systems, 43(1):115–133, Jan 2013.[103] Raffaele Gravina, Parastoo Alinia, Hassan Ghasemzadeh, and Giancarlo Fortino. Multi-

sensor fusion in body sensor networks: State-of-the-art and research challenges. Informa-

tion Fusion, 35:68–80, 2017.[104] Bluetooth SIG. Bluetooth technology website, 2017.

[105] K. Antevski, A. E. C. Redondi, and R. Pitic. A hybrid ble and wi-fi localization system

for the creation of study groups in smart libraries. In 2016 9th IFIP Wireless and MobileNetworking Conference (WMNC), pages 41–48, July 2016.

[106] Y. Peng, W. Fan, X. Dong, and X. Zhang. An iterative weighted knn (iw-knn) basedindoor localization method in bluetooth low energy (ble) environment. In 2016 Intl IEEE

Conferences on Ubiquitous Intelligence Computing, Advanced and Trusted Computing,

Scalable Computing and Communications, Cloud and Big Data Computing, Internet ofPeople, and Smart World Congress (UIC/ATC/ScalCom/CBDCom/IoP/SmartWorld),

pages 794–800, July 2016.

[107] Pavel Kriz, Filip Maly, and Tomas Kozel. Improving indoor localization using bluetoothlow energy beacons. Mobile Information Systems, 2016:11, 2016.

[108] Z. Hu, G. Huang, Y. Hu, and Z. Yang. Wi-vi fingerprint: Wifi and vision integrated finger-print for smartphone-based indoor self-localization. In 2017 IEEE International Conferenceon Image Processing (ICIP), pages 4402–4406, Sept 2017.

[109] V. Chandel, N. Ahmed, S. Arora, and A. Ghose. Inloc: An end-to-end robust indoor local-ization and routing solution using mobile phones and ble beacons. In 2016 InternationalConference on Indoor Positioning and Indoor Navigation (IPIN), pages 1–8, Oct 2016.

[110] M. Wallbaum and S. Diepolder. Benchmarking wireless lan location systems wireless lanlocation systems. In Mobile Commerce and Services, 2005. WMCS ’05. The Second IEEE

International Workshop on, pages 42 –51, july 2005.[111] D. Kotz and T. Henderson. Crawdad: A community resource for archiving wireless data

at dartmouth. Pervasive Computing, IEEE, 4(4):12 – 14, oct.-dec. 2005.

[112] Junxing Zhang, Mohammad H. Firooz, Neal Patwari, and Sneha K. Kasera. Advancing

wireless link signatures for location distinction. In Proceedings of the 14th ACM interna-tional conference on Mobile computing and networking, MobiCom ’08, pages 26–37, New

York, NY, USA, 2008. ACM.[113] Yiming Ji, S. Biaz, Shaoen Wu, and Bing Qi. Optimal sniffers deployment on wireless

indoor localization. In Computer Communications and Networks, 2007. ICCCN 2007.

Proceedings of 16th International Conference on, pages 251 –256, aug. 2007.[114] Yiming Ji. Location Determination Within Wireless Networks. VDM Publishing House

Ltd, March 2009.

102

BIBLIOGRAPHY

[115] Yiming Ji and Lei Chen. Dynamic indoor location determination: Mechanisms and robust-

ness evaluation. In Proc. Sixth Int Autonomic and Autonomous Systems (ICAS) Conf,pages 70–77, 2010.

[116] T. Vaupel, J. Seitz, F. Kiefer, S. Haimerl, and J. Thielecke. Wi-Fi positioning: System

considerations and device calibration. In International Conference on Indoor Positioningand Indoor Navigation (IPIN), pages 1–7, 2010.

[117] Jun geun Park, D. Curtis, S. Teller, and J. Ledlie. Implications of device diversity for

organic localization. In Proc. IEEE INFOCOM, pages 3182–3190, 2011.[118] Mikkel Baun and Kjrgaard. Indoor location fingerprinting with heterogeneous clients. Per-

vasive and Mobile Computing, 7(1):31 – 43, 2011.

[119] Shih-Hau Fang and Chu-Hsuan Wang. A dynamic hybrid projection approach for improvedwi-fi location fingerprinting. Vehicular Technology, IEEE Transactions on, 60(3):1037–

1044, 2011.[120] Hui Wang, A. Szabo, J. Bamberger, Dietrich Brunn, and U.D. Hanebeck. Performance

comparison of nonlinear filters for indoor wlan positioning. In Information Fusion, 2008

11th International Conference on, pages 1–7, 2008.[121] S. Ito and N. Kawaguchi. Data correction method using ideal wireless lan model in posi-

tioning system. In Personal, Indoor and Mobile Radio Communications, 2006 IEEE 17th

International Symposium on, pages 1–5, 2006.[122] C. Laoudias, M.P. Michaelides, and C.G. Panayiotou. Fault tolerant positioning using

wlan signal strength fingerprints. In Indoor Positioning and Indoor Navigation (IPIN),

2010 International Conference on, pages 1–8, 2010.[123] Bang Wang, Shengliang Zhou, Laurence T. Yang, and Yijun Mo. Indoor positioning via

subarea fingerprinting and surface fitting with received signal strength. Pervasive Mob.

Comput., 23(C):43–58, October 2015.[124] S. Sorour, Y. Lostanlen, and S. Valaee. Rss based indoor localization with limited de-

ployment load. In Global Communications Conference (GLOBECOM), 2012 IEEE, pages303–308, Dec 2012.

[125] S. Sorour, Y. Lostanlen, S. Valaee, and K. Majeed. Joint indoor localization and radio

map construction with limited deployment load. Mobile Computing, IEEE Transactionson, PP(99):1–1, 2014.

[126] W. Navidi. Statistics for Engineers and Scientists. McGraw-Hill, 2008.

[127] Bruce Craig David S Moore, George P. McCabe. Introduction to the Practice of Statistics,7th Edition. W. H. Freeman; S, 2012.

[128] S. Stavrou and SR Saunders. Review of constitutive parameters of building materials.In Antennas and Propagation, 2003.(ICAP 2003). Twelfth International Conference on(Conf. Publ. No. 491), volume 1, pages 211–215. IET, 2003.

[129] J. Jemai, R. Piesiewicz, and T. Kurner. Calibration of an indoor radio propagation pre-diction model at 2.4 GHz by measurements of the IEEE 802.11b preamble. In IEEE 61stVehicular Technology Conference (VTC 2005-Spring), volume 1, pages 111–115, 2005.

[130] G. Lui, T. Gallagher, Binghao Li, A.G. Dempster, and C. Rizos. Differences in rssi readingsmade by different wi-fi chipsets: A limitation of wlan localization. In Localization and GNSS

(ICL-GNSS), 2011 International Conference on, pages 53 –57, june 2011.[131] Shih-Hau Fang and Tsung-Nan Lin. Accurate indoor location estimation by incorporating

the importance of access points in wireless local area networks. In Proc. IEEE Global

Telecommunications Conf. GLOBECOM 2010, pages 1–5, 2010.

[132] S. Biaz, Yiming Ji, and P. Agrawal. Impact of sniffer deployment on indoor localization.In Proc. Int Collaborative Computing: Networking, Applications and Worksharing Conf,

2005.[133] George E. P. Box, William Gordon Hunter, and J. Stuart Hunter. Statistics for experi-

menters : an introduction to design, data analysis, and model building / George E. P.

Box, William G. Hunter, J. Stuart Hunter. Wiley, New York :, 1978.[134] Apple Inc. ibeacon for developers, 2017.[135] H. K. Fard, Y. Chen, and K. K. Son. Indoor positioning of mobile devices with agile

ibeacon deployment. In 2015 IEEE 28th Canadian Conference on Electrical and ComputerEngineering (CCECE), pages 275–279, May 2015.

103

BIBLIOGRAPHY

[136] Youngjune Gwon and Ravi Jain. Error characteristics and calibration-free techniques for

wireless lan-based location estimation. In Proceedings of the Second International Work-shop on Mobility Management &Amp; Wireless Access Protocols, MobiWac ’04, pages 2–9,

New York, NY, USA, 2004. ACM.

[137] Krishna Chintalapudi, Anand Padmanabha Iyer, and Venkata N. Padmanabhan. Indoorlocalization without the pain. In Proceedings of the Sixteenth Annual International Con-

ference on Mobile Computing and Networking, MobiCom ’10, pages 173–184, New York,

NY, USA, 2010. ACM.[138] Hyuk Lim, Lu-Chuan Kung, Jennifer C. Hou, and Haiyun Luo. Zero-configuration indoor

localization over ieee 802.11 wireless infrastructure. Wirel. Netw., 16(2):405–420, February

2010.[139] Anshul Rai, Krishna Kant Chintalapudi, Venkata N. Padmanabhan, and Rijurekha Sen.

Zee: Zero-effort crowdsourcing for indoor localization. In Proceedings of the 18th AnnualInternational Conference on Mobile Computing and Networking, Mobicom ’12, pages 293–

304, New York, NY, USA, 2012. ACM.

[140] Zheng Yang, Chenshu Wu, and Yunhao Liu. Locating in fingerprint space: Wireless indoorlocalization with little human intervention. In Proceedings of the 18th Annual International

Conference on Mobile Computing and Networking, Mobicom ’12, pages 269–280, New York,

NY, USA, 2012. ACM.[141] Zheng Yang, Chenshu Wu, Yunhao Liu, and Wei Xi. Will: Wireless indoor localization

without site survey. IEEE Transactions on Parallel & Distributed Systems, 24:839–848,

2013.[142] A. Nikitin, C. Laoudias, G. Chatzimilioudis, P. Karras, and D. Zeinalipour-Yazti. Indoor

localization accuracy estimation from fingerprint data. In 2017 18th IEEE International

Conference on Mobile Data Management (MDM), pages 196–205, May 2017.[143] C. G. Gavrincea, J. Baranda, and P. Henarejos. Rapid prototyping of standard-compliant

visible light communications system. IEEE Communications Magazine, 52(7):80–87, July2014.

[144] Ieee draft standard for local and metropolitan area networks - part 15.7: Short-range optical

wireless communications. IEEE P802.15.7/D3, August 2018, pages 1–412, Jan 2018.[145] H. S. Kim, D. R. Kim, S. H. Yang, Y. H. Son, and S. K. Han. An indoor visible light com-

munication positioning system using a rf carrier allocation technique. Journal of Lightwave

Technology, 31(1):134–144, Jan 2013.[146] S. De Lausnay, L. De Strycker, J. P. Goemaere, B. Nauwelaers, and N. Stevens. A sur-

vey on multiple access visible light positioning. In 2016 IEEE International Conferenceon Emerging Technologies and Innovative Business Practices for the Transformation ofSocieties (EmergiTech), pages 38–42, Aug 2016.

[147] M. S. Rahman, M. M. Haque, and Ki-Doo Kim. High precision indoor positioning usinglighting led and image sensor. In 14th International Conference on Computer and Infor-mation Technology (ICCIT 2011), pages 309–314, Dec 2011.

[148] Ye-Sheng Kuo, Pat Pannuto, Ko-Jen Hsiao, and Prabal Dutta. Luxapose: Indoor position-ing with mobile phones and visible light. In Proceedings of the 20th Annual International

Conference on Mobile Computing and Networking, MobiCom ’14, pages 447–458, NewYork, NY, USA, 2014. ACM.

[149] Y. Nakazawa, H. Makino, K. Nishimori, D. Wakatsuki, and H. Komagata. Indoor posi-

tioning using a high-speed, fish-eye lens-equipped camera in visible light communication.

In International Conference on Indoor Positioning and Indoor Navigation, pages 1–8, Oct2013.

[150] M. Yasir, S. W. Ho, and B. N. Vellambi. Indoor positioning system using visible light andaccelerometer. Journal of Lightwave Technology, 32(19):3306–3316, Oct 2014.

[151] S. De Lausnay, L. De Strycker, J. P. Goemaere, N. Stevens, and B. Nauwelaers. Optical

cdma codes for an indoor localization system using vlc. In 2014 3rd International Workshopin Optical Wireless Communications (IWOW), pages 50–54, Sept 2014.

[152] A. Cidronali, S. Maddio, G. Giorgetti, and G. Manes. Analysis and performance of a smart

antenna for 2.45-ghz single-anchor indoor positioning. Microwave Theory and Techniques,IEEE Transactions on, 58(1):21–31, 2010.

104

BIBLIOGRAPHY

[153] Yinjie Chen, Zhongli Liu, Yong Ding, and Xinwen Fu. On sampling signal strength for

localization using a directional antenna. In World of Wireless, Mobile and MultimediaNetworks (WoWMoM), 2011 IEEE International Symposium on a, pages 1–6, 2011.

[154] R. Szumny, K. Kurek, and J. Modelski. Attenuation of multipath components using di-

rectional antennas and circular polarization for indoor wireless positioning systems. InMicrowave Conference, 2007. European, pages 1680–1683, 2007.

[155] F. Penna, Jun Wang, and D. Cabric. Cooperative localization of primary users by direc-

tional antennas or antenna arrays: Challenges and design issues. In Antennas and Propa-gation (APSURSI), 2011 IEEE International Symposium on, pages 1113–1115, 2011.

105

List of Publications

The research work published in scientific papers or presented in conferences isthe following:

[1] Loizos Kanaris, Akis Kokkinis, Giancarlo Fortino, Antonio Liotta, and StavrosStavrou. Sample size determination algorithm for fingerprint-based indoor local-ization systems. Computer Networks , 101:169177, 2016.

[2] Loizos Kanaris, Akis Kokkinis, Antonio Liotta, Marios Raspopoulos, and StavrosStavrou. A binomial distribution approach for the evaluation of indoor positioningsystems. In Telecommunications (ICT), 2013 20th International Conference on,pages 14. IEEE, 2013.

[3] Loizos Kanaris, Akis Kokkinis, Antonio Liotta, and Stavros Stavrou. Com-bining smart lighting and radio fingerprinting for improved indoor localization. InNetworking, Sensing and Control (ICNSC), 2017 IEEE 14th International Confer-ence on, pages 447452. IEEE, 2017.

[4] Loizos Kanaris, Akis Kokkinis, Antonio Liotta, and Stavros Stavrou. Fus-ing bluetooth beacon data with wi-fi radiomaps for improved indoor localiza-tion.Sensors, 17(4):812, 2017.

[5] Loizos Kanaris, Akis Kokkinis, Antonio Liotta, and Stavros Stavrou. Qualityof fingerprint radiomaps for positioning systems. In Telecommunications (ICT),2017 24th International Conference on , pages 15. IEEE, 2017.

[6] Loizos Kanaris, Akis Kokkinis, Marios Raspopoulos, Antonio Liotta, and StavrosStavrou. Improving rss fingerprint-based localization using directional antennas.In Antennas and Propagation (EuCAP), 2014 8th European Conference on, pages15931597. IEEE, 2014.

[7] Akis Kokkinis, Loizos Kanaris, Marios Raspopoulos, Antonio Liotta, and StavrosStavrou. Optimizing route prior knowledge for map-aided fingerprint-based po-sitioning systems. In Antennas and Propagation (EuCAP), 2014 8th EuropeanConference on, pages 21412144. IEEE, 2014.

[8] Akis Kokkinis, Aristodemos Paphitis, Loizos Kanaris, Charalampos Sergiou,and Stavros Stavrou. Physical and network layer interconnection module for re-alistic planning of iot sensor networks. In Proceedings of the 2018 International

107

LIST OF PUBLICATIONS

Conference on Embedded Wireless Systems and Networks, pages 201202. JunctionPublishing, 2018.

[9] Akis Kokkinis, Marios Raspopoulos, Loizos Kanaris, Antonio Liotta, and StavrosStavrou. Map-aided fingerprint-based indoor positioning. In Personal Indoor andMobile Radio Com- munications (PIMRC), 2013 IEEE 24th International Sympo-sium on, pages 270274. IEEE, 2013.

[10] Marios Raspopoulos, Christos Laoudias, Loizos Kanaris, Akis Kokkinis, Chris-tos G Panayiotou, and Stavros Stavrou. 3d ray tracing for device-independentfingerprint-based positioning in wlans. 2012.

[11] Marios Raspopoulos, Christos Laoudias, Loizos Kanaris, Akis Kokkinis, Chris-tos G Panayiotou, and Stavros Stavrou. Cross device fingerprint-based positioningusing 3d ray tracing. In Wireless Communications and Mobile Computing Con-ference (IWCMC), 2012, 8th International, pages 147152. IEEE, 2012.

[12] S. Mumtaz, K. M. S. Huq, J. Rodriguez, S. Ghosh, E. E. Ugwuanyi, M.Iqbal, T. Dagiuklas, S. Stavrou, L. Kanaris, I. D. Politis, A. Lykourgiotis, T.Chrysikos, P. Nakou, and P. Georgakopoulos. Self-organization towards reducedcost and energy per bit for future emerging radio technologies - sonnet. In 2017IEEE Globecom Workshops (GC Wkshps), pages 16, Dec 2017.

108

Curriculum Vitae

Loizos Kanaris received a BSc in Aviation Science at the Hellenic AirforceAcademy in Athens, with a diploma thesis in applied aerodynamics (Computa-tional calculation of 3D flow occurring under rotary wing aircraft). After obtainingan MBA at Cyprus European University, he worked in several research projectsof the Cyprus MoD and represented the Republic of Cyprus in several technicaltelecom workgroups of international organizations, such as the COSPAS-SARSATand the IMO (COMSAR Workgroups). He is a member of the R&D department ofSigint Solutions Ltd, a telecom research performing organization, which supportsthe development of several platforms, such as TruNET wireless, a 3D, polarimet-ric, multithreading Ray Tracing Simulator. He has actively participated in variousEU and National research projects, such as WHERE (FP7), WHERE2 (FP7) andLOCME (RPF) investigating wireless positioning aspects. He has been serving asa reviewer in several scientific publications in the area of Wireless Telecom andhe has been listed as an Exemplary Reviewer in IEEE Wireless CommunicationLetters for year 2015. During his PhD (TU/e) he has been concentrating on wire-less positioning techniques and the evaluation of indoor positioning systems. Hehas published xx peer-reviewed papers and presented his work at internationalconferences. He has developed an indoor positioning evaluation tool which wasinterconnected with TruNET wireless simulator.

109

Documents

Evaluating and improving indoor positioning methods(3)WSN optimization deployment and generation of radio{maps which can be implemented in IoT and WSNs localization especially in indoor