Indoor Localization Methods Using Dead Reckoning and 3D Map Matching

  • Published on

  • View

  • Download

Embed Size (px)


<ul><li><p>Indoor Localization Methods Using Dead Reckoning and 3DMap Matching</p><p>J. Bojja &amp; M. Kirkko-Jaakkola &amp; J. Collin &amp; J. Takala</p><p>Received: 29 September 2013 /Revised: 27 November 2013 /Accepted: 2 December 2013# Springer Science+Business Media New York 2013</p><p>Abstract In order to navigate or localize in 3D space such asparking garages, we would need height information in addi-tion to 2D position. Conventionally, an altimeter is used to getthe floor level/height information.We propose a novel methodfor three-dimensional navigation and localization of a landvehicle in a multi-storey parking-garage. The solution pre-sented in this paper uses low cost gyro and odometer sensors,combined with a 3D map by means of particle filtering andcollision detection techniques to localize the vehicle in aparking garage. This eliminates the necessity of an altimeteror other additional aiding sources such as radio signalling.Altimeters have inherent dynamic influential factors such astemperature and environmental pressure affecting the altitudereadings, and for radio signals we need extra infrastructurerequirements. The proposed solution can be used without anysuch additional infrastructure devices. Other sources of infor-mation, such as WLAN signals, can be used to complementthe solution if and when available. In addition we extend thisproposedmethod to novel concept of non-stationary 3Dmaps,as moving maps, within which localization of a track-ableobject is required. We also introduce novel techniques thatenable seamless navigation solution from vehicular dead reck-oning (VDR) to pedestrian dead reckoning (PDR) and viceversa to reduce user involvement. For achieving this wecollect relevant measurements such as vehicle ignition statusand accelerometer signal variance, and user pattern recogni-tion to select appropriate dead reckoning method.</p><p>Keywords Particle filters . 3Dmapmatching . Deadreckoning . Land vehicles . Sensor fusion . Indoorenvironments</p><p>1 Introduction</p><p>Current navigation solutions employ one or several of thetechniques based on GNSS satellites and receivers, WLANdevices, inertial sensors, altimeters, and video [1]. For clear-sky out-door navigation, GNSS alone will suffice for anaccurate navigation solution. However in environments whereGNSS is totally unavailable, such as underground parkinggarages, the location information needs to be derived fromother sources such as wireless radio devices, motion sensors,and altimeters [2] [3]. Using altimeters might not be feasible,as they have inherent dynamic influential factors such astemperature and environmental pressure affecting the altitudereadings [4].</p><p>In this paper, we study 3Dmap-matching in parking garages,a scenario different from the common map-matching problemin various senses. Firstly, in parking garages, GNSS cannot berelied on due to the heavy attenuation of satellite signals whenpenetrating concrete structures; therefore, one has to resort tousing on-board motion sensors such as the odometer of thevehicle. Secondly, vehicle heading is less constrained than onroads and streets, which poses additional challenges when agyroscope is used for heading estimation. Thirdly, positioninginmulti-storey parking garages requires the use of 3Dmaps andknowledge on the altitude of the vehicle.</p><p>This paper proposes a method for achieving low cost andeffective solution to such GNSS denied indoor multi-storeyparking garage navigation. Many existing devices in the mar-ket, such as smartphones, are equipped with gyros, acceler-ometers, and all modern land vehicles (e.g., cars) have odom-eters. The speed information of the vehicle is obtained by asmart phone via a Bluetooth on-board diagnostics (OBD)reader supporting protocol version II connected to the vehicle.Given the nonholonomic constraints of vehicle motion and theinitial location of the vehicle with respect to the target indoor,these sensors readings and a detailed 3D map are sufficient to</p><p>J. Bojja (*) :M. Kirkko-Jaakkola : J. Collin : J. TakalaDepartment of Pervasive Computing, Tampere University ofTechnology, Tampere, Finlande-mail:</p><p>J Sign Process SystDOI 10.1007/s11265-013-0865-9</p></li><li><p>obtain an indoor 3D positioning solution on a smartphone.This paper extends our preliminary results reported in [5].</p><p>A 3D model as depicted in Fig. 1(a), representing thestructural details of a real-world multi-storey parking garageshown in Fig. 1(b), is used as a 3D map and motion constraintin the solution which is based on particle filtering. In the filter,each particle is modelled as a separate 3D vehicle objectwhich has approximately the same horizontal and verticaldimensions as a true vehicle. This 3D vehicle object in itselfcan be used as a map within which localization is required.This enables us to localize the passengers inside movingvehicles such as trains, boats and busses, for example. Sincethese 3D maps of vehicles are moving with respect to theEarth-fixed frame wherein GNSS location is traditionallyexpressed, the term moving maps is used in this article.</p><p>The rest of the paper is organized as follows. Section 2describes the related work, Sections 3 and 4 address theparticle filtering and collision detection methods upon whichthe proposed navigation algorithm relies. Section 5 describesthe technically extended novel concepts of moving maps andseamless navigation solution. Section 6 describes the mea-surement and experimental setup for testing, and demonstrat-ing the localization approach, using real-world sensor dataobtained by driving a car in parking garages. Finally, Section 7concludes the paper.</p><p>2 Related Work</p><p>Map-matching has been studied for decades, with the firstimplementations estimating the position of a vehicle along aknown route [6]; an extensive description of the most com-mon map-matching algorithms is given in [7].</p><p>Many solutions to the 3D indoor positioning problem havebeen proposed in the literature. Wagner et al. [8] used cascad-ed Kalman filters and road link matching, for positioningvehicles in parking garages. In [9], Fouque et al. describes ageneric solution to multi-hypothesis map-matching usingglobal positioning on tightly integrated 3D navigable road</p><p>maps, formalized in a general Bayesian framework. Pintoet al. [10] proposes a 3D map based approach to pinpoint arobot pose, by using 3D map of the surrounding environmentand data acquired by a tilting laser range finder.</p><p>Nowadays, a popular approach is to use a particle filter(PF); they are known to be well suited for positioning prob-lems [11]. In [12], Fairfield et al. proposed a submap-based3D simultaneous localization and mapping (SLAM) known assegmented SLAM or SegSLAM, and used Cave Crawlerrobotic vehicle for obtaining speed and heading sample data,using on-board wheel based odometer and laser range mea-surement system respectively. The basic idea of segmentationis to circumvent the scale limitation inherent in SLAM. Theyapply Rao-Blackwellized particle filter for SLAM and extendit to allow particles to transition between sub maps. Kmmerleet al. [13] proposed an autonomous driving in a complexmulti-storey parking garage, using a modified vehicleequipped with multi laser range finders, a high-performanceinertial measurement unit and GPS receivers. They used a PFfor localizing (SLAM) the car in and within a 3D map of theparking garage environment by only using the laser range datafor generating the map and using the inertial data for the PFand localization. They have used 1000 particles and limitedthe speed of the vehicle to 10 km/h and an update of data at200Hz. However, the map information cannot be used as anefficient motion constraint in SLAM because the map is oneof the unknowns. Leppkoski et al. [14] proposed a pedestriandead reckoning solution for indoor pedestrian navigation withdetailed indoor maps as a motion constraint, inertial sensors asthe primary source of information, and radio signals asassisting signals. This study showed that a very detailed 2Dmap, including even bookshelves, significantly improved thePF navigation solution in a 2D space.</p><p>3D models and map matching has been studied earlier. In[15] Andreja et al. proposed a mobile robot self-localization incomplex indoor environments using monocular vision and 3Dmodel of the environment. The captured noisy and complexvideo frames are processed for real time 2D image segmenta-tion and line extraction by using Canny edge detector and</p><p>Figure 1 Multi-storey parking garage: a 3D model and b the real world garage.</p><p>J Sign Process Syst</p></li><li><p>random window randomized Hough transform (RWRHT).The extracted edges are then matched with the 3D model linesegments to localize the robot. The odometry device in therobot is used to estimate the robots pose and use it as theestimated camera pose for rendering 3D model. Then edgematching procedure is applied to adjust this assumed camerapose that is rendering the 3D scene to match with the edges ofthe captured 2D image. Ascher et al. [3] proposed a multi floorindoor pedestrian navigation, where stairs, elevators and lad-ders are used for transitioning from floor to floor. This methoduses barometer, magnetometer/compass and inertial measure-ment units as sensor data input to the algorithm and particlefilter system based on bootstrap particle filter implementation[16]. Here a 3D model of the scenario (building) is used forcompensating for the errors of the sensors such as magneticinterference causing heading errors and the accumulated er-rors of accelerometer and gyro. There are several solutionproposals for automatic map generation, such as for 2D,2.5D, and 3D maps [1921, 2832].</p><p>Our approach is to use a detailed 3D structural map of theparking garage depicted in Fig. 2(a), as a motion constraint tonavigate in 3D space using on-board motion sensors. In ourPF, we test each particle for collisions in a novel way, takinginto account the dimensions of the vehicle which improves theaccuracy of the map-matching. Instead of using altimeters orany other means for height estimation we use the ramp struc-tures of the 3D models, eliminating the necessity for external</p><p>radio navigation updates, which are a common solution as in[17] when no map information is used. We also extend theapplicability of this method for navigation in trains, buildingswith elevators etc. by introducing moving map concept. Innavigation related literature very often the research is concen-trated on a specific motion (for example driving or walking) orenvironment (indoor or outdoor) and the algorithms for thetransition phase receive little attention. In this paper, wepropose a solution for automatic switching from VDR toPDR, enabling a seamless navigation solution.</p><p>3 Particle Filtering</p><p>Particle filtering is an approximation of the Bayesian filterwhere the posterior distribution p (xn |y1,,n,x0), with xndenoting the state vector at time step n and y1,,n being themeasurements, is characterized by a cloud of random samples,called particles, instead of, e.g., the moments of the distribu-tion. The foremost benefit of this representation is the abilityto operate on arbitrary distributions, thus making it possible toestimate, e.g., multimodal distributions which often causedivergence of Kalman-type and other filters that assumeGaussian distributions. Particle filtering is a Monte Carlomethod and both its performance and computational complex-ity depend on the number of particles used. We use a PFvariant called the bootstrap filter [18] where the so-calledimportance distribution is chosen to be the transitional priordistribution.</p><p>Suppose we have N particles x i with nonnegative weightswi, i =1,,N . Each particle is a state vector containing thequantities that are to be estimated; in this study, the i th particleis a 41 vector</p><p>xi Hi</p><p>Ei</p><p>N i</p><p>Ui</p><p>264</p><p>375 1</p><p>whereH is the heading angle, E ,N , andU denote East, North,and vertical coordinates, respectively. Particle filtering con-sists of two basic steps, i.e., prediction and updating. In theprediction phase, we draw particles from the transitional dis-tribution. Assuming that nonholonomic constraints hold, theexpected value of this distribution can be expressed as</p><p>E xin xin1</p><p> Hin1 n t</p><p>Ein1 cos Hin Vn t</p><p>N in1sin Hin</p><p> Vn tU in1 tan Pin</p><p> Vn t</p><p>2664</p><p>3775 2</p><p>where E is the expected value operator; P is the inclinationangle of the surface on which the particle is located; Vn andn are the measured speed and angular rate at the n</p><p>th time</p><p>Figure 2 a Detailed structural model of a single floor parking garage andb virtual sensor spheres attached to a vehicle model at different placesused in collision detection.</p><p>J Sign Process Syst</p></li><li><p>step, respectively; and t is the measurement interval. Thecovariance of the transitional distribution is determined basedon e.g., the error characteristics of the motion sensors.</p><p>In the update step, the weights of the particles are modifiedaccording to the likelihood of a measurement given the statevector. In the case of the bootstrap filter, the update is doneaccording to the simple proportion</p><p>winp ynxin</p><p> win1: 3</p><p>Due to the proportionality relation, the weights need to benormalized to sum to unity after updating. This way, it isstraightforward to estimate the mean of the posterior distribu-tion as the weighted average of the particles.</p><p>In this paper, we use the 3D map as a source of measure-ment updates according to the likelihood function</p><p>p yn xin</p><p> 0 if the particle hit a wall1 otherwise</p><p>: 4</p><p>In other words, particles that collide are discarded. In thisstudy, the map update likelihood is computed by modellingeach particle as a vehicle with physical dimensions and mo-tion constraints, instead of a freely moving point mass. Thissignificantly improves the accuracy of the map-matching be-cause we can detect if one of the corners of the vehicle istouching a wall although the centre of mass is not. Section 4discusses more on this kind of modelling and about collisiondetection.</p><p>If other measurements are available; they can be incorpo-rated into the estimation process by means of additionalupdate steps with appropriate likelihood functions. It is obvi-ous that discarding colliding particles leads to a situationwhere only a small fraction of the N particles are actuallyused for the state estimation; such a cloud of particles isobviously not a good approximation of a probability distribu-tion and also causes a waste of computational resources ifzero-weighted particles are propagated. This problem can beavoided by resampling the set of particles; in this procedure, anew set of N particles is drawn from the discrete probabilitydistribution defined by the old particles and their respectiveweights, such that the newly obtained set of particles representthe same distribution as the old one, but with a full number ofalive particles. A block diagram of the flow of the signalprocessing in the PF system is shown in Fig. 3.</p><p>4 Collision Detection</p><p>Collision detection is defined in this context as the ability tocomputationally detect if two or more objects...</p></li></ul>