Bearing-only SLAM in Indoor EnvironmentsUsing a Modified Particle Filter

N. M. Kwok and G. DissanayakeARC Center of Excellence in Autonomous Systems (CAS)

Faculty of EngineeringUniversity of Technology SydneyBroadway, NSW, 2007, Australiangai.kwok,gdissa@eng.uts.edu.au

AbstractThe implementation of a particle filter (PF)for vision-based simultaneous localisation andmapping (SLAtvI) for a mobile robot in an un-structured indoor environment is presented inthis paper. Variations to standard PF are pro-posed to remedy the sample impoverishmentproblem in bearing-only SLAM. A CCD cameramounted on the robot is used as the measuringdevice and image quality is incorporated intodata association, PF update and map manage-ment. A passive path control strategy to main-tain the accuracy of the SLAM process is alsoillustrated. Experimental results from an im-plementation using real-life data acquired froma Pioneer robot are included to demonstratethe effectiveness of our approach.

1 IntroductionIn mobile robot applications, it is a fundamental require-ment that the robot should be able to know its positionwithin its operating environment. This can be achievedby using odometry measurements and referring to a mapof the environment. However, odometry feedback tendsto accumulate errors in time, therefore other measuringdevices are needed. A detail and precise map is alsonot always available and we have to build the map asthe robot operates. Therefore, we have to solve the si-multaneous localisation and mapping (SLAM) problem.Solutions to SLAM [Dissanayake et. al., 2001] in indoorenvironment use a laser range-and- bearing scanner andput the problem in the extended Kalman filter (EKF)framework. The laser scanner is accurate but is heavyand expensive. The EKF operates satisfactorily whenthe system is not severely non-linear and non-Gaussian.If this does not hold true, one has to employ particlefilters (PF) [Gordon et. al., 1993] for a better represen-tation of the probability density function.

Apart from using a laser scanner as the measuring de-vice, cameras are popular alternatives because of their

reduced cost, weight and power consumption. Relatedwork can be found in [Davison, 2002] where a pan-tiltcamera is used for SLAM and also uses the EKF. Workon a combination of laser and vision is reported in [Perezet. al., 1999] and [Arras et. al., 2000]. Use of vision andodometry is reported in [Crowley, 1992], where land-marks are known and listed in a database. Similarly,in [Mufioz and Gonzalez, 1998], pre-stored landmarksare also used in localising a mobile robot using a sin-gle image. Another approach by pre-recording imagesequences is found in [Jeon and Kim, 1999] but this re-stricts the operating area of the robot. Although theuse of vision has its many advantages, detection and se-lection of landmarks from image sequences is challeng-ing [Livatino and Madsen, 1999]. On the other hand,some research work for outdoor SLAM can be found in[Fitzgibbons and Nobet, 2001].

The PF is based on the Monte Carlo simulation tech-nique and uses samples to represent the system probabil-ity distribution function (PDF). This filter out-performsthe EKF for non-linear systems. However, the PF islimited by its need for a large number of samples torepresent the PDFs making it computational expensive.The PF also suffers from the so-called sample impov-erishment problem that samples tend to converge to aconfined region in the solution space, making state esti-mations trapped in local optimums. Methods to remedythe problems can be found in [Gordon et. al., 1993].

This paper presents research on the use of a fixed cam-era mounted on a mobile robot for simultaneous locali-sation and mapping in an unstructured indoor environ-ment. Landmarks are extracted from edges in the sceneand stored as subsequent matching templates while therobot is moving.

To ensure that the PF operates over an extendedperiod of time, we tackle the sample impoverishmentproblem by introducing techniques from genetic algo-rithms (GA). Similarities between PF and GA have beenpointed out in [Higuchi, 1997]. Further reference isalso found in [Bienvenue et. al., 2001] where multi-

1

objective estimation is treated. Variations to the PFusing GA techniques are proposed in [Deutscher et. al.,2001] where enhancements to the PF is achieved witha crossover operator. Another way of modifying PFsis found in [Meier and Ade, 1999] where the PDF istruncated resulting in an extended tail. The PDF istruncated that allows for dynamic inclusion of newly ob-served landmarks. Here, we use the BLX-o operatorwhich fills the undefined supports of the PDF and ex-pends it [Herrera et. al.. 1998].

To improve the robustness of SLAl\L we incorporatethe image quality [Wang and Bovik, 2002] for data as-sociation in addition to the usual nearest-neighbour val-idation test. Image quality is used in the update step ofthe PF as motivated by the data association method in[Dezert and Bar-Shalom, 1993]. Map management as in[Dissanayake et. al.. 2002] is also used where we removelandmarks by an index combined from image quality anddetected edge height. \Ye also consider the effect of robotpath that affects the SLA::..rperformance. Related workcan be found in [Bianco and Zelinsky, 2000] where navi-gation guidance is treated.

The rest of this paper is arranged as follows. In section2, we briefly review the SLAM problem. Then, the PF isintroduced in section 3. In section 4, modifications to theparticle filter will be presented. Data association usingimages will be discussed in section 5 followed by the mapmanagement in section 6. In section 7, we will propose apassive path control strategy. Experiment results will bepresented in section 8 and the discussion and conclusionin section 9.

2 SLAM ProblemGiven a mobile robot deployed in its operating environ-ment, we denote its position at time zero as the originof the world co-ordinate system. That is

where :1:1), Yv is the location in Cartesian co-ordinates ando; is the heading with reference to the x-direction

Fig.l depicts the system definitions. When a control11. consisting of speed and turn rate, is issued: the robotmoves according to a transition equation .

[

xv(k + 1)] [X"(k) + vtcos(rPv(k) + ,(k)) + Vx ]yv(k + 1) yv(k) + vtsm(rPv(k) + ,(k)) + vy

rPv(k + 1) rPv(k) + ,(k)t + v"(2)

where v is the velocity command, "( is the turn rate. t isthe time step and l.'xy6 are noise terms lumping in theeffects of control response. wheel slip, etc" the noises arecharacterised by N(O, axy¢)

While moving, the robot makes measurements to land-marks in the operating environment. The observation is

landmark 1 \

XI1'Yll \ r landmark 2

Xf2'Y12

range and -.Ibearing

to landmark 1,r.e.

\'--- range and

bearing tolandmark 2,

(282

robot orientation 11\

robot location x ~' y ~

origin of worldco-ordinate

Figure 1: System definition

ziven bvb •

(Xfi-X,,)

(j = arctan - 01) + VeYfi - Y,·

(3)

where :1: fi, Yfi is the location of a stationary landmark,VII is the measurement noise N(O, ae)

The SLAM problem is to estimate the aggregatedrobot and landmark locations. in the form of a state vee-tOL provided with the measurement, The state vector isgiven as

3 Particle Filter

(1)

The particle filter (PF) is based on the Bayesian estima-tion framework. Information about the system is embed-ded in a probability distribution function (PDF). Using alarge number of samples (particles) as in the Monte Carlomethods, the PDF is represented and evolves accordingto the system transition equation. Upon reception of ameasurement. the PDF is updated and the estimationprocess repeats.

At time zero, we have the system PDF given by p(xo).The system then evolves according to the transitionequation, giving p(i:k+llxk), When a measurement isavailable, a likelihood of obtaining that measurement canbe found as p( Zk+l Ii:k+l), Finally, we combine the PDFaccording to the Bayes' rule and get a recursive equation

Details of theoretical development can be found in [Gor-don et. al.. 1993]. Implementation of the particle filtercan be described as follows.

1. randomly initialise particles. xi

2

2. in every time step, propagate particles through thetransition equation, xi <- xi

3. calculate the likelihood (weight) from the measure-ment, p(Zi[xi)

4. re-sample (select) particles according to theirweights, xj <- Xi

The implementation of a PF includes propagating theparticles through the transition equation. However, inthe SLAM problem, the landmarks are stationary andthe landmark particles! will soon collapse onto a singlepoint, which is called sample impoverishment.

4 Modified Particle FilterThere are several research works on the similarity be-tween particle filtering and genetic algorithms (see intro-duction). We will make use of GA techniques to modifythe PF implementation.

4.1 CrossoverWe will adopt the BLX-o: operator [Herrera et. al., 1998]to move the landmark particles in each iteration. Theprocedure is given below.

1. select 2 particles Xi, xj, with equally likely proba-bility

2. calculate the distance dij between the 2 particles inCartesian co-ordinate

3. move the particles according to the product of the0: parameter and the distance, giving xij = xij ±(0: x dij)

4. further move the particles by a uniform random pa-rameter r E [01], giving Xij = r x Xi + (1 - r) x xj

Here, we choose parameter 0: = 0.5. This choice resultsin spreading of particles and helps in reduce the sampleimpoverishment problem by introducing new potentialsolutions to the estimation. Referring to [Gordon, 1994],this corresponds to jittering and prior editing.

To minimise the effect of the sample redistribution onthe shape of the PDF, we limit the crossover using asmall crossover probability (Pc = 0.1) and this appearsto work well in practice. We are currently in the pro-cess of developing a more rigorous algorithm to find anappropriate value for Pc-

Landmark particles far away from the robot havesmaller bearing error than those particles near the robotprovided they have the same perpendicular displacementfrom the bearing measurement. The result is that there-sampling process favours particles far away and makesthe estimation biased. Using the same method as above,we again move the particles as follows.

INote that a particle represents the system vector, weviolate the terminology here for the ease of description

1. separate landmark particles into 2 groups withequally likely probability

2. for 1 group, rotate the particles by 1800 centred atthe expectation value of the 2 groups, that is

When we rotate the particles at the expectation value,this value is un-altered. The rotation by 1800 does notinvolve any scaling or translation and this retains thecovariance between the particles. The result is thatwe have redistributed the particles leaving their statis-tics un-altered. This reduces far away particles to there-sampling process and consequently reduces the bias.Fig.2 through Fig.4 shows typical landmark particle his-tograms (PDF) before and after rotate and move.

4.2 Batch Update

Each update in the particle filter involves re-sampling,thus increases the possibility of sample impoverishment.We try to minimise this possibility by performing batchupdate by accumulating a series of measurements beforeeach update.

4.3 Measurement Quality

The likelihood function represents the probability of ob-taining the real-life measurement given the expectedrobot and landmark locations. When using a camera asa measuring device, we can incorporate the image qual-ity into the calculation of the likelihood function. Weuse the following equation to calculate the likelyhood.

(7)

where ei is the expected bearing measurement for a par-ticle and Q E [01] is the quality of the landmark image(see section5). When the image quality is poor, mostparticles will be retained through the re-sampling pro-cess. That is, we take a rather conservative approachwhen measurement is not sure then reduce the effect ofupdating.

5 Data Association

Data association maintains the correspondence betweena measurement and a landmark. It is crucial for theoperation of any estimation. In this work, we will use ahybrid association that the nearest neighbour and imagequality are combined. We also use hysteresis thresholdsto reduce ambiguities.

3

0

0

0

0

0

I

i0

20 ! I

i III10

0I ." L-

Figure 2: Landmark PDF before rotate

,II, ,II00

Figure 3: Landmark PDF after rotate and before move

35

Figure 4: Landmark PDF after move

5.1 Nearest Neighbour

This is a classical approach that determines the differ-ence between the real-life and the expected measure-ments. The result is thresholded against some confidencelevel in the X2 sense. That is, if

then declare the measurement from a particular land-mark. Where v = (}- {j is the innovation, a is theerror covariance between the robot and the landmark,")'1 ;:::; 3.8 for 95% confidence with 1 degree of freedom(bearing-only measurement).Otherwise, if

(9)

then declare the measurement from a new landmark.Where ")'2 ;:::; 10.8 for 99.9% confidence.

5.2 Image qualityA camera is used in our work as the measuring deviceproviding the bearing-only measurement for SLAM. Im-age quality then plays a critical role in matching imagepatches for data association. We follow the criteria pro-posed in [Wang and Bovik, 2002] using correlation, con-trast and intensity characteristics as a metric betweensuccessful image frames I, and Ij. Image I, is the patchof image in the current frame and Ij is the previous im-age frame stored during PF iterations. The image qual-ity Q E [01[is defined as

(10)

where aij is the correlation between the images I, andt., a«, aj are the variances of the 2 images, 7 and J arethe mean values respectively

To reduce ambiguity, hystersis threshold is used. IfQ > ")'3 then declare as matched landmark, otherwise, ifQ < "/4 then declare as new landmarks. Where

")'3 = mean(Q) + 0.5 x std(Q)

")'4 = mean(Q)

(11a)(11b)

5.3 Edge DetectionLandmarks from the image frame are detected by findingtheir edges using the sobel mask. Threshold is also usedthat we accept edges when their edge heights are abovethe mean values from an image frame. Point landmarksmay be detected, however, they are liable to occlusions.We also use spatial separation to resolve ambiguities thatwe only extract edges from a horizontal strip of the mid-dle of the image. This also speeds up the image pro-cessing speed. Fig.5 and Fig.6 show typical scenarios inedge detection, initialising new landmarks and trackinglandmarks.

5.4 Overall Data Association

(8)We summarise the overall data association below. First,we pick measurements from nearest neighbour. Second,

4

Figure 5: Previous frame with detected edges

Figure 6: Current frame, initialised (left) and detectededges

compute the image quality and pick the matched edges.Among the measurements, pick the one with

as matched measurement and tracked landmark. Other-wise, initialize new landmarks by picking the measure-ments corresponding to

6 Map ManagementIt has been shown in [Dissanayake et. al., 2002] that itis possible to trim the system state by removing someof landmarks that is not measured and is of low cer-tainty. In this work, we make use of edge height andimage quality as the criteria to remove landmarks. Weset a maximum number of landmarks, m = 15, in our im-plementation. When the number of landmarks is morethan m, then remove those tracked landmarks with asmallest rank. That is, remove the landmark using

where eg is the stored edge height during tracking, Q isthe store image quality

When a new landmark is found, it is initialised by spa-tial separation that a certain angle span, Os, among thetracked landmarks has to be met. It is a reasonableassumption that landmarks are distributed uniformlywithin an indoor environment where the robot operates.Employing our strategy, landmarks with high edges andhigh image qualities will remain being tracked.

7 Path ControlWith bearing-only SLAM, appropriate path followingstrategy should be derived for the success of SLAM. Inthis case, we are inferring the 2-dimensional robot andlandmark position from the l-dimensional bearing-onlymeasurement. The convergence to the true estimationdepends on the bearing sustained between the robot andlandmarks. A ±90° measurement maximises the rate ofconvergence. However, our camera has a field-of-view of90° and maximum convergence rate cannot be obtained.

V'-Iepropose here, to steer the robot in a circular pathwith periodic panning back of the robot. This strategysimulates an extended field-of-view, and the estimationresult (see experiment section) outperforms that of purecircular path. By panning the back robot, we may limitthe robot uncertainty while it is moving. If this uncer-tainty is bounded, newly found landmarks will have alower uncertainty and be kept tracked. Bounding therobot uncertainty also enhances the success rate in dataassociation.

(12)

8 Experiment ResultsExperiment results from range-and-bearing SLAM us-ing a laser scanner are used as the reference to verifythe vision-based bearing-only experiments. The laserscanner is relatively accurate and with laser reflectorsas landmarks ensures the robustness of data associationand estimation, and is reasonably justified as the refer-ence. However, because of difficulties in detecting laserreflectors as landmark in image frames, it is not possibleto make a direct comparison for the landmark locations.As a compromise, we show the laser scans in the figuresthat illustrate vision-based experiments (note that thelaser scans are not used in the estimation) to imposean indication of the operating environment. We havekept 15 landmarks in the estimation and that amountsto 33 states including the robot. In the experiment, weuse 10000 particles and obtain extended PF operationthrough 400 time steps.

8.1 Circular PathIn this experiment, we move the robot in a continuouscircular path for about 2 rounds. Fig.7(a) shows a snap-shot during the estimation using the laser range-and-

(13)

(14)

5

Loop 200

-2

-4

-6

-6 -4 -2

(a) laser range-and-bearing

(b) vision bearing-only

MobilaRobot Error Plot -3Cf

!~F~; Io 50 100 150 200 250 300 350 400!~:r~~4o 50 100 150 200 250 300 350 400

(E~~c 50 100 150 200 250 300 350 400

Time Steps

(c) robot position error

Figure 7: Experiment results for Circle Path

bearing measurements. Fig.7(b) shows the correspond-ing snap-shot for the vision-based bearing-only SLAM.It can be observed that landmark estimations are notas good as the range-and-bearing, however, this is ex-pected as more information is available in the formercase. Fig.7(c) shows a plot of the robot position estima-tion error with the corresponding covariance 3er bound.Because of extended close-loop time, errors tend to ac-cumulate and it is clear that the estimate becomes in-consistent as time is increased.

8.2 Pan-back PathWe basically move the robot in a circular path, but wepan back the robot for about 90° when the robot movesabout a quarter of a circle. Fig.8(a) shows the laserrange-and-bearing experiment result. The vision-basedbearing-only experiment snap-shot is shown in Fig.8(b).The result using the laser scanner is quite similar to theprevious case. This is due to the fact that the laser scan-ner is relatively accurate and has a large field-of-view(180°), thus making the estimation rather independentof the path. The vision-based case shows a better per-formance than the circular path case. This is due to thefact that we periodically pan the robot back and limitsthe estimation uncertainties. The landmark estimationis inferior to the laser range-and-bearing case, which isagain expected. Fig.8(c) shows the robot position esti-mation error. A better performance is observed than thecircle path case. However, the robot heading estimationshows periodic fluctuations, we suspect this is due to er-rors in the process model that become apparent whenthe robot turns rapidly during pan back.

9 Discussion and ConclusionA vision-based bearing-only SLAtvI is implemented us-ing a particle filter. Preliminary experiment results showthat extended operation of the particle filter is possi-ble by manipulating the particle samples by adoptingthe crossover technique from genetic algorithms. Imagequalities are incorporated in filter operation, data associ-ation and map management. The results also show thateffective path planning is crucial to obtaining acceptableestimation quality, particularly as the field of view of thecamera used is limited to 90°.

AcknowledgementThis work is supported by the ARC Centre of Excellenceprogramme, funded by the Australian Research Council(ARC) and the New South Wales State Government.

6

6 Loop 200

1,

,~. ]If'

-2

-4

(a) laser range-and- bearing

-2

-4

-6 -4 -2

(b) vision bearing-only

Mobile Robot Error Plol - 30!~:L~~~~~o 50 100 150 200 250 300 350 400

!£======:;Jo 50 100 150 200 250 300 350 400

i:[;\t;~(v;-200 50 100 150 200 250 300 350 400

TimeSleps

(c) robot position error

Figure 8: Experiment results for Pan-back Path

References[Arras et. al., 2000] Arras K. 0" Tomatis N. and Siegwart

R., Muliiserisor On-the-Fly Localisation Using Laser andVision. Proc. 2000 IEEE/RSJ Intl. Conf. on IntelligentRobots and Systems, 462-467, 2000.

[Bianco and Zelinsky, 2000] Bianco G. and Zelinsky A.,Real-time Analysis of the Robustness of the Naviga-tion Strategy of a Visually Guided Mobile Robot. Proc.Intl. Conf. on Intelligent Autonomous Systems, IAS2000,Venice, July 2000.

[Bienvenue et. al., 2001] Bienvenue A., Joannides M., Be-rard J., Fontenas E. and Francois 0., Niching in MonteCarlo Filtering Algorithms. Proc. 5th Intl. Conf. on Arti-ficial Evolution, 19-30, Creusot, October 2001.

[Crowley, 1992] Crowley J. L., Position Estimation for a Mo-bile Robot Using Vision and Odometry. Proc. 1992 IEEEIntl, Conf. on Robotics and Automation, 2588-2593, Nice,May 1992.

[Davison,2002] Davison A. J., SLAM With a Single Camera.SLAM/CML Workshop at ICRA 2002, 11-15, May 2002.

[Deutscher et. al., 2001] Deutscher J., Davison A. and Reid1., Automatic Partitioning of High Dimensional SearchSpaces Associated with Articulated Body Motion Capture.Proc. 2001 IEEE Compt. Society Conf. on Comput. Visionand Pattern Recognition, 669-676, December 2001.

[Dezert and Bar-Shalom, 1993] Dezert J. and Bar-ShalomY., Joint Probobiliiic Data Association for AutonomousNavigation. IEEE Trans. on Aerospace and Electronic Sys-tems, 29(4):1275-1285, October 1993.

[Dissanayake et. aI., 2001] Dissanayake G., Newman P.,Clark S., Durrant-Whyte H. F. and Csorba M., A Solu-tion to the Simultaneous Localisation and Map Building(SLAM) problem. IEEE Trans. on Robotics and Automa-tion, 17(3):229-241, June 2001.

[Dissanayake et. aI., 2002] Dissanayake G., Williams S. B.,Durrant-Whyte H. F. and Bailey T., Map Management forEfficient Simultaneous Localisation and Mapping (SLAM).Autonomous Robots, 12:267-286, 2002.

[Fitzgibbons and Nobet, 2001] Fitzgibbons T. and Nobet E.,Application of Vision in Simultaneous Localization andMapping. Proc. 2001 Australian Conf. on Robotics andAutomation, 121-127, Sydney, November 2001.

[Gordon et. aI., 1993] Gordon N. J., Salmond D. J. andSmith A. F. M., Novel Approach to Nonlinear/Non-Gaussian Bayesian State Estimation. lEE Proceedings F,Radar and Signal Processing, 140(2):107-113, April 1993.

[Gordon, 1994] Gordon N. J., Non-linear / Non-GaussianFiltering and the Bootstrap Filter. IEE Colloquium onNon-linear Filters, 1-6, May 1994.

[Herrera et. al., 1998] Herrera F., Lozano M. and VerdegayJ. L., Tackling Real-coded Genetic Algorithms: Operatorsand Tools for Behavioural Analysis. Artificial IntelligenceReview, 12(4):265-319, 1998.

[Higuchi, 1997] Higuchi T., Monte Carlo Filter Using theGenetic Algorithm Operators. Journal of Statis. Comput.Simul., (59):1-23, 1997.

[Jean and Kim, 1999] Jean S. H. and Kim B. K., Monocular-based Position Determination for Indoor Navigation ofMobile Robots. Proc. lASTED IntI. Conf. on Control andApplications, 408-413, Banff, July 1999.

7

[Livatino and Madsen. 1999] Livatino S. and Madsen C. B.,Optimization of Robot Self-Localisation Accuracy by Au-tomatic Visual-Landmark Selection. Proc. SCIA99 11thScandinavian Conf. on Image Analysis, 501-506, Kanger-lussnag, June 1999.

[:\Ieier and Ade, 1999] Xlcier E. B. and Ade F .. TrackingCars in Range Images Using the Condensation Algorithm.Proc. 1999 IEEEjIEEJ JJSAI Intl. Conf. on IntelligentTransportation Systems, 129-134, October 1999.

[Mufioz and Gonzalez, 1998] Mufioz A. ,J. and Gonzalez J.,Two-Domensional Landmark-based Position Estimationfrom a Single Image. Proc. 1998 IEEE IntI. Conf. onRobotics and Automation, 3709-3714, Lcuven, :\[ay 1998.

[Perez et. aI., 1999] Perez J. A., Castellanos J. A., MontielJ. M. M., Neira J. and Tardos J. D., Continuous MobileRobot Localisation: Vision vs. Laser. Proc. 1999 IEEE IntI.Conf. on Robotics and Automation, 2917-2923. Detroit,:\[ay 1999.

[Wang and Bovik. 2002] Wang Z. and Bovik A. C., A Uni-versal Image Quality Index. IEEE Singal ProcessingLeters, 9(3):81-84, March 2002.

8

nan KODOnCS OCf\.UlUUli111UU /"\.I)I)Ul-IC1UVl1 Hi".

Australasian Conference on Robotics and Automation 2003

Foreword

Welcome to ACRA 2003.

ACRA, the Australasian Conference on Robotics and Automation, is the annual conference for the Australian Robotics andAutomation Assodation, and Australia and New Zealand's leading forum for work in all areas of robotics and automation. ACRAattracts researchers and practitioners from around Australia and New Zealand, and an increasing number of roboticists fromoverseas. The work presented at ACRA 2003 is highly relevant to current interests in international robotics research, including hottopics such as autonomous mapping and humanoid robots. ACRA 2003 also has a definite Australian robotics flavour, with anemphasis on outdoor robotics and applications in the primary industries that continue to form the basis of the Australian and NewZealand economies in the 21st century.

We have endeavoured to ensure a balanced programme: maintaining a high standard of quality, while being inclusive of the up andcoming students and junior researchers in the Australasian robotics community. The Programme Committee accepted 55 full papersfor presentation and publication after a review process in which full manuscripts were reviewed by at least two committee members.The accepted papers were drawn from 64 papers submitted in total. In keeping with recent trends, the primary medium for theproceedings is CD-ROM, and we will be maintaining the proceedings online at the ARAA website: www.araa.asn.au. In addition to thetechnical paper programme, the programme includes a number of interesting live robotics demonstrations, the Annual GeneralMeeting for the ARAA, and plenty of social opportunities for Australasian and international robotics researchers to network anddevelop cross-institutional links.

We hope you enjoy your time at the conference and that you come away knowing a little bit more, having met and discussed roboticsissues with a colleague, having attended a paper session, or having had a quite technical debate over dinner. To all, we warmlywelcome you to ACRA 2003, and to Brisbane.

Jonathan Roberts & Gordon WyethCo-chairs ACRABrisbane, December 2003

on behalf of the Programme Committee.

http://www.araa.asn.au/acra/acra2003/html/foreword.html 22/04/2004

Australasian Conference on Robotics and Automation 2003

Table of Contents

ApplicationsRobots for pre-orientation and interaction of toddlers and preschoolers who are blindB Bartlett, V Estivill-Castro, S Seymon & A Tourky

Driver assistance: contemporary road safetyAndrew Dankers, Luke Fletcher, Lars Peters son & Alexander Zelinsky

Karlsruhe surgical robotics researchJorg Raczkowsky. Sascha Dauber. Dirk Engel. Harald Hoppe, Werner Korb, Oliver Schorr. StefanHassfeld & Heinz Worn

DNA analY§l~ using a portable robotic instrumentDaksh Sadarangani, Bruce Macdonald, Allen G Rodrigo & David Saul

Walking RobotsAutomatic gait optimisation for quadruped robotsMin Sub Kim & William Uther

Control of a heavyweight biped robot in the frontal planeJoe Cronin, Richard Willgoss & Robin Ford

Cerebellar joint compensation for a humanoid robotDamien Kee & Gordon Wyeth

Improved joint control using a genetic algorithm for a humanoid robotJonathan Roberts. Damien Kee & Gordon Wyeth

Multi-Robot TeamsAchieving cooperation in a distributed multi-robot teamMark M Chang & Gordon Wyeth

Production rules as chromosomes of GA for robotic swarm applicationsKai Wing Tang & Ray A Jarvis

Behavioural approach for multi-robot explorationHaye Lau

Graphical simulation and visualisation tool for a distributed robot programming environmentFelix-Etienne Trepanier & Bruce A MacDonald

Classifying an opponent'.§J;l§hayiQill JQJQQQtsoccerDavid Ball and Gordon Wyeth

VisionCreating composite images for estimation the effectiveness of mobile robot coverage algorithmsSylvia C Wong, Lee Middleton & Bruce A MacDonald

Representation and learning of visual information for pose recognitionDavid Prasser & Gordon Wyeth

Fusion of multimodal visual cues for model-based object trackingGeoffrey Taylor & Lindsay Kleeman

An ordered list approach to real-time line detection based on the hough transform

http://www.araa.asn.aulacralacra2003/html/contents.html

Video 1, Video 2, Video~, Video~, Video 5Video 1, Video 2

22/04/2004

Australian Robotics & Automation Association Inc.

Chris Madden & Robert Mahony

Red is the new black - or is it ?David Austin & Nick Bames

Sensors IWind sensor and robotic model wasp developmentDavid Harvey, Tien-Fu Lu & Michael Keller

Na'ive physics for effective odour localisationGideon Kowadlo & R Andrew Russell

Chemical source location and the RoboMole projectR Andrew Russell

Object location and recognition using whisker sensorsR Andrew Russell & Jaury Adi Wijaya

Vision and ProgrammingA survey of robot programming systemsGeoffrey Biggs & Bruce MacDonald

Sensing for visual homingKane Usher, Matthew Dunbabin, Peter Corke & Peter Ridley

performance of temporal filters for optical flow estimation in mobile robot corridor centring and visualodometryChris McCarthy & Nick Bames

Techniques for impro'{iQ9-vL§]Q!lqI}QJQcornotion on the~Q!lY_AHlQrob9tMichael J Quinland, Stephan K Chalup & Richard H Middleton

NavigationHOMER - a high speed robot for indoor explorationNathan Wilson, Matt Rozyn & Michael Andrews

Combining wavefront propagation and possibility theory for autonomous navigation in an indoorenvironmentCindy Leung & Adel AI-Jumaily

High precision GPS guidance of mobile RobotsRichard Willgoss, Vivian Rosenfeld & John Billingsley

Short-safe compromise patch for mobile robot navigation in a dynamic unknown environmentSardjono Trihatmo & R A Jervis

Covert path planning for autonomous robot navigation in known environmentsMohamed Marzouqui & Ray Jarvis

On robotic path planning using rapidly exploring random treesZoltan Deak Jnr & Ray Jarvis

Outdoor RobotsSubmarine dynamic modellingPeter Ridley, Julien Fontan & Peter Corke

Submarine automatic controlPeter Ridley, Julien Fontan & Peter Corke

T~rr9iJ] aidedl,mc!er:water navig<ltion - a deeper insight into generic monte carlo localizationAlexander Bachmann & Stefan B Williams

Helicopter automation using a low-cost sensing systemGregg Buskey, Jonathan Roberts, Peter Corke & Gordon Wyeth

A helicopter named Dolly - Behavioural cloning for autonomous helicopter controlGregg Buskey, Jonathan Roberts & Gordon Wyeth

ManipulatorsOn-line singularity avoidance, collision detection and task level programming utilizing a world model

http://www.araa.asn.au/acra/acra2003/htmllcontents.html

Page 2 of3

22/04/2004

Australian Robotics & Automation Association Inc. Page 201'3

Per Cederberg, Magnus Olsson & Bunnar Bolmsjo

The effect of ill-conditioned inertial matrix on controlling robot manipulatorYueshi Shen & Roy Featherstone

New ways of generate better configurations using geometric featuresAntonio Benitex & Daniel Vallejo

A chattering-free variable structure controller for tracking of robotic manipulatorsTri V MNguyen, Q P Ha & Hung T Nguyen

Robots in AgricultureControl of industrial robots for meat processing applicationsZeng Li, Peter Ring, Kym MacRae & Andrew Hinsch

A new approach to detech the cutting position for a robotic beef carcass scribing systemZeng Li & Andrew Hinsch

Machine vision system for counting macaQamia nutsMark Dunn & John Billingsley

Modular decentralized control of fruit picking redundant manipulatorTimothy R Vitior, Richard A Willgoss & Tomonari Furukawa

Localisation and MappingA Hybrid approach to finding cycles in hybrid mapsMargaret E Jefferies, Wenrong Weng, Jesse T Baker, Michael C Cosgrove & Michael Mayo

Three-Dimensional robotic mappingIan Mahon & Stefan B Williams

Line-based SMC SLAM method in environment with polygonal obstaclesDavid C K Yuen & Bruce A MacDonald

Bearing-only SLAM in indoor environments using a modified particle filterN M Kwok & G Dissanayake

Six Dof D-SLAMLee Ling (Sharon) Ong, Matthew Ridley, Jong-Hyuk Kim, Eric Nettleton & Salah SUkkarieh

Hippocampal models for simultaneous localisation and mapping on an autonomous robotMichael Milford & Gordon Wyeth

Sensors IITracking people with networks of heterogeneous sensorsAlex Brooks & Stefan Williams

Active monocular fixation using the log-polar sensorAlbert Yeung & Nick Barnes

Uncertainty of line segments extracted from static SICK PLS laser scansAlbert Diosi & Lindsay Kleeman

Proximity Identification and tracking for mining automationGarry Einicke

http://www.araa.asn.au/acra/acra2003/html/contents.html 22/0412004