Reference scan matching for global self-localization

Robotics and Autonomous Systems 40 (2002) 99–110

Reference scan matching for global self-localization

Joachim Weber∗, Lutz Franken, Klaus-Werner Jörg, Ewald von PuttkamerComputer Science Department, University of Kaiserslautern, P.O. Box 3049, D-67653 Kaiserslautern, Germany

Abstract

Especially in dynamic environments a key feature concerning the robustness of mobile robot navigation is the capability ofglobal self-localization. This term denotes a robot’s ability to generate and evaluate position hypotheses independently frominitial estimates, by these means providing the capacity to correct position errors of arbitrary scale. In the self-localization framedescribed in this article, the feature-based anchor point relation (APR) scan matching algorithm provides for each new laser scanseveral possible alignments within a set of reference scans. The generation of alignments depends on similarities between thecurrent scan and the reference scans, only, and is unconstrained by earlier estimates. The resulting multiple position hypothesesare tracked and evaluated in a hybrid topological/metric world model by a Bayesian approach. This probabilistic techniqueis especially designed to integrate position information from different sources, e.g. laserscanners, computer vision, etc.© 2002 Elsevier Science B.V. All rights reserved.

Keywords: Global self-localization; Mobile robot navigation; Scan matching; Service robots; Probabilistic reasoning

1. Introduction

Robustness of mobile robot self-localization is de-termined by the navigation algorithm’s capacities torecover from erroneous position estimates. In contrastto industrial applications, where human and machineworking spaces are usually separated and the degreeof environmental dynamics is controllable, servicerobots have to deal with populated, unprepared,possibly mutable and generally unpredictable envi-ronments. Due to temporary environmental changes,parts of the known world might become unrecogniz-able for the robot’s external sensors. Thus, the robothas to rely on internal position sensors (e.g. odometryor inertial systems), accumulating an uncorrected po-sition error while travelling without external revision.Tracked self-localization techniques (e.g.[1,5]) are

∗ Corresponding author. Tel.:+49-631-205-2656;fax: +49-631-205-4409.E-mail address: [email protected] (J. Weber).

designed torefine an existing position estimate by ex-ternal sensor observations within a limited tolerance,only. Hence, they are unable to recover from accumu-lated position errors beyond that tolerance. In orderto overcome these deficiencies,global localizationtechniques (e.g.[4,6,9,11]) create and by probabilisticmeans evaluate a multitude of position hypotheses inparallel. If the robot loses the correct position in anunrecognizable region, global self-localization willsooner or later award the correct position with themaximum likelihood after having returned to recog-nizable areas, again. However, since global techniquesface a significantly bigger search space than trackedlocalization methods, they are generally less precisein order to still fit real time requirements.

Closely connected to the question of how to de-sign global localization algorithms is the preferredenvironment representation (world model), whichitself is determined by the available sensors andthe intended robot application. Basically, two dif-ferent types of environment representations can be

0921-8890/02/$ – see front matter © 2002 Elsevier Science B.V. All rights reserved.PII: S0921-8890(02)00235-X

100 J. Weber et al. / Robotics and Autonomous Systems 40 (2002) 99–110

identified: metric and topological world models. Ametric environment map, storing obstacles and fea-tures in a unique coordinate system, is a straight-forward and easy-to-implement approach whenrangefinders like laserscanners and ultrasonic sensorsare used. However, metric maps are not well suited tointegrate non-metric information as required by manypattern matchers (e.g. the neural image classifier in-troduced by Wichert[16]). Furthermore, metricallyconsistent map building is a non-trivial problem, cur-rently being under intensive research (e.g.[8,12]).Global self-localization techniques based on metricworld models are[2,4,6].

Topological representations, instead, model the en-vironment as graphs of nodes (distinctive places) andconnecting edges (pathways). The basic advantagesof topological world models compared to metric onesare:

• simplified map building, since only topological, notprecise metric consistency has to be achieved,

• path planning is reduced to direct graph search,• non-metric and metric sensor data is uniformly

treated as attributes to nodes and/or edges, whichsimplifies sensor fusion.

However, topological world models suffer from thepoor environment resolution, which is insufficient forany other purpose than navigating from one region toanother. Examples for global self-localization basedon topological maps are[9,11].

Whenever metric and non-metric localizationtechniques are mixed, the use of ahybrid topologi-cal/metric world model is adequate to benefit from theadvantages of both environment representations, whileat the same time avoiding their inherent restrictions.

In the CAROL project (CameraBasedAdaptiveRobot Navigation and Learning), which servesas a framework for the development of globalself-localization techniques and navigation architec-tures, various metric or appearance-based algorithmsfor processing laserscanner and vision data have beendeveloped and implemented. Consequently, the natu-ral base to fuse these different techniques is a hybridworld model, which is explained in detail inSection3. One of the fundamental localization methods, theanchor point relation (APR) scan matching approachdescribed inSection 2is by itself some sort of hy-brid, since the idea of recognizing a certain region by

a stored sample scan is essentially topological. TheAPR recognition process, however, results in metricalignment information of current and reference scan.The probabilistic method that tracks and evaluatesmultiple position hypotheses, integrating results fromdifferent sensors is treated inSection 4.

2. APR scan matching

In the past, various scan matching techniques havebeen proposed for different kinds of applications.These approaches can be separated into two maincategories:

Feature-based matching techniques either alignextracted geometric primitives or raw range readingswith an existing structural description of the environ-ment. These matching techniques represent an effi-cient, reliable and popular class of self-localizationmethods for polygonal environments (e.g.[1,3]).However, feature-based techniques suffer from thepoor assortment of practically usable geometric prim-itives, thus mainly line segments and derived land-marks, e.g. corners and door openings are used.

In recent years a second category of more flexi-ble raw data matching techniques without explicitgeometric interpretation have been developed (e.g.[4,5,10]). Unfortunately, most of these scan match-ing approaches like[5,10] are designed to refine anexisting position estimate, hence are not suitable forglobal self-localization. On the other hand, the globalapproach presented by Crowley et al.[4] directly in-tegrates the matching and position tracking techniquein a strict metric world model. Especially, the closeintegration of matching and tracking technique renderthe fusion of localization information from differentsensors difficult.

The APR scan matching technique[13] treated inthis section supports global localization by the realtime search of matches for an actual laser scan in agiven set of reference scans. The algorithm’s output ineach processing cycle aremultiple, weighted match-ing hypotheses. The quality of every single matchingalternative is computed depending on the degree ofcongruence between the raw data of current and refer-ence scan, with respect to the hypothetical alignmentof both scans. This quality measure is especially im-portant in probabilistic self-localization architectures.

J. Weber et al. / Robotics and Autonomous Systems 40 (2002) 99–110 101

Since there is no explicit geometrical interpretationby the extraction of structural primitives, APR avoidsthe typical information loss associated with the con-centration on lines and clusters, only. However, inorder to be real time capable, the alignment of twolaser scans is computed depending on the geometricalpatterns formed by extracted feature points (anchorpoints—APs). In environments which do not providea sufficient number of these features, alignments can-not be determined.

2.1. AP extraction

In our current implementation, three types of APsare used:jump edge, angle and virtual edge APs.The extraction steps are rather technical and thereforemainly skipped in this article. Details concerning theAPR AP extraction can be found in[13].

Jump edge anchors andangle anchors both refer toobject edges which intersect the scanning plane. Jumpedges refer to convex edges where only one of the en-closing object faces is visible from the scanning point,and are detected basically by a jump of consecutiverange readings. Angle anchors refer to convex or con-cave edges when both respective object faces are visi-ble. Briefly described, they can be detected where therange values on a scanned object are continuos but the

Fig. 1. Laserscan and extracted APs (left) and resulting fully connected AP graph (right).

object’s surface angles are discontinuous. In order toconcentrate on objects of a minimum size for local-ization purposes, small objects (e.g. chairs, wastebas-kets, etc.) are removed from further consideration byexecuting a segmentation step, first. This segmenta-tion step finds large connected parts in the range scan,and jump edge and angle anchors are searched in thesesegments only.

Virtual edge anchors are intersection points of theelongations of large straight objects’ surfaces (e.g.walls, lockers, etc.). In the most simple case, the cor-ners of a room are extracted as virtual edge anchors,even if the corner itself is hidden by some object.Please note that the intersections are found withoutexplicit line extraction. Instead, geometrical statisticsis used to determine the large objects’ positions andorientations, which offers higher robustness againstpartial occlusion and noise effects. The basis for therecognition of a scan in the APR matching is its ge-ometrical AP pattern, which is mathematically givenby the fully connected AP graph (seeFig. 1).

2.2. Preselection

Usually, two scans of the same scenery will resultin two disjoint sets of APsA andA′ due to three mainsources of error:


• occlusion of significant environment parts by ob-jects which were not present at creation time ofthe corresponding reference scan (dynamic environ-ments),

• fluctuations in position extraction of APs due tomeasurement errors and aliasing effects,

• only partial overlap of the two corresponding 180◦scans due to different scan positions and orienta-tions.

The matching problem is thus reduced to the iden-tification of maximum matching subgraphs in the setof all reference graphs. In order to avoid a full searchin the reference set, APR performs a preselection oflikely candidates using therelationship database. Thedatabase search tries to identify reference graphs witha similar spatial AP configuration by statistic means. Inoffice environments usually a high number of similardistances between single pairs of APs will occur (dueto standardized furniture, identical door sizes, etc.).However, only the same or mirror-symmetrical con-figuration of objects will produce a similar, fully con-nected graph and thus a high number of correspondingedge lengths.

The relationship database is constructed fromall edges of all reference graphs. Each edgeei =(pa, pb, di) of a reference graphGn is representedin the database as a tuple (n, ei). Both APspa andpb are nodes inGn and di is the euclidian distancebetween APpa andpb. The database is implementedas an index table sorted according to distancesdi .

The preselection of probable matching candidatesin the reference set for graphG′ of the current scan isdone by a search for distancesdi in the relationshipdatabase for allei ∈ G′. For each edge entry (m, ej ) inthe database, which corresponds in length to anei ∈G′, a countercm for the respective reference graphGm

is incremented. Thus, after the completed search of all

Fig. 2. Alignment of two AP graphs using a correspondence matrix.

edges ofG′ in the database, a countercm contains thenumber of edge length correspondences of referencegraphGm andG′. The preselection qualityqm for eachreference graphGm is then computed as

qm = c2m

kact · km

wherekact andkm are the total numbers of edges inG′andGm, respectively. Only reference scans with a highpreselection valueqm are chosen for further process-ing: the alignment step. Theqm measure’s construc-tion prevents ‘big’ reference graphs with many edgesof becoming ‘more attractive’ only due to a resultinghigh number of accidental correspondences.

2.3. Alignment

The alignment step aims at finding a coordinatetransformation to align one AP graphG1 (e.g. a refer-ence graph) with another graphG2 (e.g. the AP graphof the actual laser scan). This not only provides thepossibility to check the validity of the match by somemeans of raw data correlation; in the case of a correctmatch it directly delivers the actual robot position inthe reference graph’s local coordinate system.

Each length correspondence of two edgese1 ∈ G1ande2 ∈ G2 provides two possibilities for the align-ment of both graphs. Fore1 = (pa, pb, d1) ande2 =(pv, pw, d2) with d1 ≈ d2, the combinations (pa , pv)and (pb, pw) might represent the same APs in the twographs, or (pa , pw) and (pb, pv), or none of both ifthe correspondence is just accidental.

The alignment decision is computed statistically bycreating ann×m AP correspondence matrix (Fig. 2),wheren and m are the total numbers of APs inG1and G2, respectively. For each of the above men-tioned four possible AP combinations (pi , pj ) per edge


correspondence, the respective matrix element (i, j) isincremented by 1. The matrix thus counts the possibleAP identifications indicated by the edge correspon-dences as a sort of two-dimensional histogram.

In the example ofFig. 2, G2 is a true subgraph ofthe largerG1, but with a different AP enumeration. InG1 the edges (1, 2) and (2, 4) have the same length(i.e. the same length than (2, 3) inG2), all other edgeshave different lengths. The resulting correspondencematrix achieves a maximum in line 2, column 3, whichidentifies the pair AP 2 ofG1 and AP 3 ofG2 as themost likely AP correspondence.

APR now uses the edge correspondences whichhave contributed to this maximum entry to compute anequal number of geometric alignment hypotheses. Allcorrect contributions to that peak will produce similarhypotheses, erroneous contributions are eliminated bya clustering step. The remaining transformations areunified by computing the center of gravity to providea more precise alignment.

As can be seen in the example ofFig. 2, the‘noise-to-signal-distance’ in the correspondence ma-trix is not satisfactory for small total number of APsor poor overlap ofG1 and G2. In order to elimi-nate coincidental edge correspondences, like (2, 4)in G1 with (2, 3) in G2, APR determines the mostlikely rotational difference betweenG1 andG2 froma matching angle histogram. This histogram resultsfrom a frequency analysis of the angle differencesbetween all pairs of corresponding edges inG1 andG2. All edge correspondences not belonging to thehistogram peak are ruled out from further considera-tion in the AP correspondence matrix. InFig. 2, thefalse edge correspondence of (2, 4) inG1 with (2, 3)in G2 will be eliminated because three correct edgecorrespondences indicate a rotational difference of30◦ between the two graphs, whereas the false edgepair suggests an angle of−45◦ (modulo 180◦).

Although the matrix technique is not recommend-able for general graph matching problems with highnode numbers, it is very efficient and, by its statisticalnature, highly robust against AP extraction errors fortypical graph sizes in the APR application scenario.

2.4. Evaluation

APR performs the alignment step of the actualscan for a selected number of reference scans with

the highestqm values. The evaluation step providesa quality measureem for the use in a probabilisticself-localization architecture.

Using 180◦ laser scans bears the problem of differ-ent perspectives; generally, scan A contains environ-mental information which scan B does not, and viceversa. Additionally, range data acquired from differentpositions, but representing the same part of the envi-ronment, cannot be correlated directly.

This problem is solved by creatingsynthetic scans;firstly, the relative translation of scans A and B ischecked to determine which scan contains the otherscan’s sensing position and consequently has a better‘overview’. From this scan’s range data an artificialscan for the other scan’s (front scan) position and ori-entation is computed. This synthetic range image canbe directly correlated with the front scan to determinethe degree of congruence achieved after alignment.

2.5. Experimental results

Fig. 3 shows a set of 30 reference scans of our labenvironment, fitted into a global coordinate system. Aseries of 1405 scans recorded during a second, inde-pendent test run was matched with this reference set.The SICK scanner is mounted at a height of approxi-mately 25 cm, thus the algorithm has to cope with acluttered field of ‘vision’ due to chair and table legs,waste paper baskets, etc. For each scan matching cy-cle, five reference graphs with the highest preselectionvalues were checked for correspondences with thecurrent laser scan. All scan matches wereglobal, i.e.no a priori available position information was used toreduce search space.

In Fig. 3, the APR matching results are marked bytriangles in the map. For each of the 1405 matchingcycles only the alignment result with the highest eval-uation ratingem is displayed. Correct position esti-mates are colored gray, wrong estimates are coloredblack. A line between gray triangles indicates a tracksection without matching result, i.e. at these positionsthe evaluation valueem of the best alignment resultwas below a certain threshold. In total, 613 correctglobal position estimates and 1 wrong estimate weregenerated, which shows a high hit/error ratio. Aver-age scan processing time including all preprocessingsteps was below 15 ms with a maximum of 37 ms percycle on a 233 MHz Pentium PC.


Fig. 3. Global APR scan matching results. Correct matches are indicated by gray triangles, the false match is marked by a black triangle.

If only the reference scan with the highest preselec-tion valueqm is aligned with the current scan, the num-ber of correct matches decreases to 374 with no falsematching. If all reference scans are checked in eachcycle, 831 correct matches and 1 false are producedwith a still acceptable computing time of 57.4 ms (av-erage) and 119 ms (max.).

Please note that the procedure of transforming localalignment results into a global coordinate frame cor-responds to global self-localization in ametric naviga-tion scheme. At the same time, APR performs well inidentifying thetopologically correct reference scans.Environmental ambiguities are handled by providingmultiple different, reasonably quantified hypotheses.This is especially of interest for the probabilisticself-localization technique described inSection 4.

3. World model

In industrial applications, a mobile robot oftenrequires its precise position in a unique coordinatesystem. This is especially necessary for precise trackfollowing on defined pathways and global movementcoordination of multiple robots. In these applica-tions, the environment can generally be presumed as

cooperative. So are employees in an automated pro-duction plant instructed to keep clear of a movingrobot’s path to avoid work flow disturbances. In theservice domain, robots have to operate in unpreparedand populated areas, e.g. office rooms, supermarketsand museums. Persons in the robot’s working space,e.g. shoppers in a supermarket, are not necessarilycooperative. The use of predefined pathways andglobal movement coordination of a robot fleet is tooinflexible to cope with that kind of unpredictable anduncooperative environment. Hence, the necessity ofa global metric coordinate system in many serviceapplications is questionable. However,local metricprecision is still often required, e.g. for cleaningpath planning, station docking, etc., which rules outa purely topological approach. Furthermore, in or-der to profit from the geometrical localization dataof the APR alignment results or other metric local-ization methods, the world model has to have somemetric component.

Consequently, we argue to use a topological graphas the backbone of environment representation, whileat the same time establishing local metric coordinatesystems in the nodes’ regions (Fig. 4). This preservesthe advantages of local precision, simple data fu-sion and easy path planning, while eliminating the


Fig. 4. Local coordinate systems in the hybrid world model andbidirectional storage of sensor data.

necessity for the expensive and difficult establishmentof global metric map correctness. In our scheme, theapproximate relative positions and orientations of thelocal coordinate systems of two topological neigh-bors are stored in the connecting arc. Thus, a method,which is tracking the robot’s position from one lo-cal system to the next, has to deal with small errors,only. Lacking a global reference system, the basicdifficulty during map building is the decision whento close the connection between two already exist-ing nodes and to determine their local coordinatesystems’ relative position and orientation. This prob-lem is solved by APR’s and other global localizationtechniques’ abilities to recognize distinctive places ofthe environment.

There is no explicit global reference system definedin the map, although such a common coordinate sys-tem can be assembled from the particular metric neigh-bor information. However, due to the summation of amultitude of minor errors between single nodes, thissystem will be contorted and ‘torn’ in those places,where its successive construction closes a topologicalcycle. During map construction, however, even suchan inconsistent global coordinate system can providea raw idea about the likelihood of a closed cycle, in-dicated by APR or another technique.

In topological maps, sensor data usually is storedin the nodes’ data structures. Due to the exclusive useof directed sensors in the CAROL project (180◦ laserscanner and camera), we chose to place sensor databidirectionally in the arcs’ ends, instead. Thus, therobot ‘knows’ a node’s appearance for each registeredarc in both directions of that particular pathway.

4. Probabilistic position tracking

In classictracked self-localization approaches (e.g.[1,5,10]) one robot position estimate is observed, only,and new sensor data is merely used to proceed andoptimize that position belief in space. These Kalmanfilter or correlation techniques generally provide highprecision in metric world models. However, sincebeing unimodal, these techniques are able to com-pensate limited position errors, only. Thus, a positionerror accumulated in a temporarily unrecognizableenvironment part might be beyond the algorithm’serror recovery capabilities, causing the robot to per-manently loose its correct position estimate.

In contrast,global self-localization has to bemul-timodal, which denotes the ability to track multiplehypotheses in parallel. A common approach (e.g. in[2,6,9,11]) is to model position belief as probabilitydistribution over the target environment. Bayes ruleis used to fuse current sensor information and lastposition belief in update cycles. The decision howto discretize the position probability distribution isessentially influenced by the underlying world model.In topological representations (e.g.[9,11]) probabilityvalues are usually computed fortransitions betweennodes, rather than for the nodes themselves. This al-lows the consideration of the robot’s rough orientation.In the metric Markov approach introduced by Burgardet al. [2] the volume of the robot’s three-dimensionalconfiguration space (with dimensionsx, y and orienta-tion φ with respect to the global coordinate system) isdiscretized into fine-grain three-dimensional grids. Inthe update cycles the probability value for each gridcube is computed depending on the match of currentrange data and map contents. Due to its large compu-tational expense, this method has been superseded bya more efficient sample-based method presented byFox et al.[6].

As mentioned earlier, the purely topologicalapproaches do not provide the desired resolution ofposition estimate. The metric techniques mentioned(including the eigenspace method by Crowley et al.[4]), on the other hand, presuppose the existence of aconsistent metric map, since there is no high-level pro-cessing of sensor data which would allow a perspec-tive-independent recognition of distinctive places.Instead, they rely on the map’s property to providefor each hypothetical robot position and orientation a


set of range data which is directly comparable to thecurrent sensor readings. An inconsistently assembledmap is a handicap for this ability in significant parts ofthe environment. Furthermore, none of the above men-tioned metric concepts is really prepared to integrateinformation apart from range data. However, a greatadvantage of all of these approaches is their principalability to cope with arbitrary environmental structures.

The position tracking mechanism developed inthe CAROL project, like the topological methodsof Simmons and Koenig[11], and Kortenkamp andWeymouth[9] directly supports the fusion of sensorinformation from different sources, thus providingmore robustness by a higher degree of environmentinformation. On the other hand, our technique pro-vides precise position estimates like the metric tech-niques[2,4,6]. However, due to the renunciation ofa unique coordinate system, these estimates refer tolocal coordinates in the nodes’ regions only. Due tothe renunciation of a consistent and complete metricmap, the algorithm would not be able to provide theenvironment’s appearance for every possible positionand orientation. Instead, our tracking algorithm hasto rely on the applied localization techniques’ (e.g.APR) abilities to generate independent position hy-potheses. In contrast to the global metric approaches[2,4,6], this principally limits the possible target en-vironments to those whose characteristics can behandled (i.e. recognized) by the sum of the appliedpattern matchers. In order to increase respective flex-ibility, self-organizing pattern matchers which adaptthemselves to the environmental characteristics havebeen implemented (e.g. the neural image classifierproposed by Wichert[16]) and developed (e.g. neuralscan classification[14]) in the CAROL project.

The following description of the probabilistic track-ing technique is an outline of the basic idea. The mech-anism has been formally described in[15].

4.1. Position tracking

With respect to a node’s local coordinate sys-tem, we define an associatedlocal three-dimensionalconfiguration space for the possible robot positions(translationx, y and orientationφ) in this node’s re-gion. The union of all local configuration spaces formsthe non-consistentsuper space, in analogy to the as-sembled, inconsistent global coordinate system. We

furthermore assume the existence of at least oneprimary source of localization information, at leastone tracking sensor and possibly furthervalidationsources.

Primary sources, like APR, are localization tech-niques which generate global position estimates in-dependently of earlier assumptions. To stay with theAPR example, all alignment alternatives for the cur-rent laserscan exceeding a certain quality thresholdare used to generate or confirm so-calledexplicitposition hypotheses. These data structures discretizethe robot’s current position probability distribution inthe super space. Each position hypothesisi is definedby a volumeVi in a certain local configuration spaceand a probability valueλi , indicating the likelihoodof the current robot position actually being insideVi (Fig. 5). The volumeVi is given by a centerci

with an uncertainty region defined by a radiusri oftranslational tolerance and the angle tolerance∆i .Theλi ’s have to sum up to 1 together with the valueof a special ‘rest’ hypothesis. This value representsthe likelihood that the actual robot position isnotrepresented by one of the explicit hypotheses.

First step after receiving new valid APR resultsis the check whether their geometrical position esti-matesconfirm existing position hypotheses, i.e. if anyof APR’s position alternatives lies in one of theVi ’s.Since the robot might currently be heading out of onenode’s context into another, all position hypotheseshave to be checked. This is done in the super spacewith careful attention to its known contortions. If anexisting position hypothesis is confirmed by an APRresult, it changes to the configuration space of therespective reference scan’s ‘home’ node, the centerci

Fig. 5. Local configuration space and position hypothesis.


is set to the APR position estimate, the uncertainty re-gion is reset to a small default value, and probabilityλi

is rewarded according to the confirming APR result’squality. APR results which do not confirm already ex-isting hypotheses are used to create new ones in therespective local configuration spaces with an initiallylow λi-likelihood. The probabilities of unconfirmedposition hypotheses are “punished” and hypotheseswith λi values below a certain threshold are deleted.

At this point the question is still open how to mod-ify the probability of the ‘rest’ hypothesis. Since thisvalue should reflect the system’s ‘skepticism’ againstthe explicit position hypotheses, it is rewarded or pun-ished periodically, depending on whether or not exist-ing hypotheses have been confirmed by sensor data.

Validation data sources, in contrast to primarysources, are inadequate to create new hypotheses. Thismight, e.g. result from high computational require-ments, which do not allow a global search (e.g. theiterative IDC algorithm[10] which requires a goodstarting point). However, they are able to provideuseful information to confirm or reject existing posi-

Fig. 6. Topological map of test environment with APR reference scans; test course for self-localization test.

tion hypotheses and to refine aci ’s local precision.Confirmations from validation sources are treatedexactly like those from primary sources.

Tracking sensors, e.g. odometry or inertial systems,are used to move all position hypotheses (i.e. to mod-ify their ci ’s) according to the measured relative move-ment since creation or last confirmation, respectively.The tracking sensors are important to bridge passagesin unrecognizable environment parts. Since they donot provide any information about a hypothesis’ like-lihood, theλi proportions are unaffected by trackingsensors. However, since their data is generally erro-neous, the growing position uncertainty is reflected byexpanding theVi volumes.

4.2. Experimental results

In the experimental setup, APR serves as the pri-mary, odometry as tracking and a neural image clas-sification technique[16] as validation source. Theimage classification is performed by a growing neuralgas net[7] structure created by a sample set of envi-


Fig. 7. Position hypotheses (black triangles) after first APR matching cycle (a), 32 cycles (b), 75 cycles (c) and 106 cycles (d).


ronment images. In this example, the image classifierreprents the sample set through 150 reference vectors(neurons), enabling the algorithm to distinguish be-tween 150 different classes of visual scenes.Fig. 6shows the topological backbone of the environmentmap and the 53 APR reference scans stored in thegraph. They are assembled in an inconsistent globalcoordinate system (size approximately 50 m× 70 m)for visualization purposes. Additionally to the refer-ence scans, image scene classes have been registeredin the topological graph during map building. The pathindicated in the scan map is the robot’s test course.

In order to give an impression of the creation andtracking of position hypotheses, the hypothesesci arevisualized inFig. 7for the first 30 m of the robot’s testcourse. The hypothesesci are indicated by triangles,the strongest hypothesis is marked by a circle. Fromthe first APR results on, the correct position estimatedominates the other explicit hypotheses. (InFig. 7λbindicates the probability of the best,λ0 the probabil-ity of the ‘rest’ hypothesis.) Please note that no scansare available in the middle of two corridors. In theseregions all doors had been closed and since no APRfeatures had been detectable, no reference scans wereinserted into the topological graph. Although APRdoes not work in these track sections, the correct posi-tion hypothesis receives confirmations from the imageclassification. Thus, it remains on a high relative prob-ability level, reinforcing the correct position estimateeven in monotonous environment parts until reachingthe final position.

References

[1] K.O. Arras, S.J. Vestli, Hybrid, high-precision localisationfor the mail distributing mobile robot system MOPS, in:Proceedings of the IEEE International Conference on Roboticsand Automation (ICRA-98), Leuven, Belgium, 1998.

[2] W. Burgard, D. Fox, D. Hennig, T. Schmidt, Positiontracking with position probability grids, in: Proceedings ofthe First Euromicro Workshop on Advanced Mobile Robots(EUROBOT), Kaiserslautern, Germany, 1996.

[3] I.J. Cox, Blanche: An experiment in guidance and navigationof an autonomous robot vehicle, IEEE Transactions onRobotics and Automation 7 (2) (1991) 198–204.

[4] J.L. Crowley, F. Wallner, B. Schiele, Position estimates usingprincipal components of range data, in: Proceedings of theIEEE International Conference on Robotics and Automation,Leuven, Belgium, 1998.

[5] T. Edlinger, G. Weiß, Exploration, Navigation and Self-localization in an Autonomous Mobile Robot, AutonomeMobile Systeme (AMS), Karlsruhe, Germany, 1995.

[6] D. Fox, W. Burgard, F. Dellaert, S. Thrun, Monte Carlolocalization: Efficient position estimation for mobile robots,in: Proceedings of the 16th National Conference on ArtificialIntelligence (AAAI-99), Orlando, FL, 1999.

[7] B. Fritzke, A growing neural gas network learns topologies,in: Advances in Neural Information Processing 7 (NIPS1994), MIT Press, Cambridge, MA, 1995.

[8] J.-S. Gutmann, K. Konolige, Incremental mapping of largecyclic environments, in: Proceedings of the InternationalSymposium on Computational Intelligence in Robotics andAutomation (CIRA), Monterey, CA, 1999.

[9] D. Kortenkamp, T. Weymouth, Topological mapping formobile robots using a combination of sonar and visionsensing, in: Proceedings of the 12th National Conferenceon Artificial Intelligence (AAAI-94), Seattle, WA, 1994,pp. 979–984.

[10] F. Lu, E.E. Milios, Robot pose estimation in unknownenvironments by matching 2D range scans, in: Proceedingsof the IEEE Conference on Computer Vision and PatternRecognition (CVPR), Seattle, WA, 1994.

[11] R. Simmons, S. Koenig, Probabilistic navigation in partiallyobservable environments, in: Proceedings of the InternationalJoint Conference on Artificial Intelligence (IJCAI-95),Montreal, Canada, 1995, pp. 1080–1087.

[12] S. Thrun, W. Burgard, D. Fox, A probabilistic approachto concurrent mapping and localization for mobile robots,Machine Learning and Autonomous Robots (Joint Issue) 31(5) (1998) 29–53.

[13] J. Weber, K.-W. Jörg, E. von Puttkamer, APR—global scanmatching using anchor point relationships, in: Proceedings ofthe Intelligent Autonomous Systems (IAS-6), Venice, Italy,2000.

[14] J. Weber, A. Spieß, K.-W. Jörg, E. von Puttkamer,Selbstorganisierende Klassifikation von 2D-Laserscans zurNavigation eines AMR, Autonome Mobile Systeme (AMS),Karlsruhe, Germany, 2000.

[15] J. Weber, L. Franken, K.-W. Jörg, K. Schmitt, E.von Puttkamer, An integrative framework for globalself-localization, in: Proceedings of the IEEE Conference onMultisensor Fusion and Integration for Intelligent Systems(MFI), Baden-Baden, Germany, 2001.

[16] G. von Wichert, Selforganizing visual perception for mobilerobot navigation, in: Proceedings of the First EuromicroWorkshop on Advanced Mobile Robots (EUROBOT),Kaiserslautern, Germany, 1996.

Joachim Weber was born in Mayen,Germany in 1968. He received hisDipl.-Inform. degree from the Universityof Kaiserslautern’s Department of Com-puter Science in 1996. Since then he worksas a research assistant at the Robotics andProcess Control Research Group at theUniversity of Kaiserslautern. His researchinterests include sensor data processing,robot learning and self-localization.


Lutz Franken was born in Olpe, Ger-many in 1972. He studied computer sci-ence at the University of Kaiserslauternwith a focus on AI and robotics. In2000, he received his Dipl.-Inform. degreefrom the Department of Computer Science.Currently, he works as an IT consul-tant for BERATA GmbH in Böblingen,Germany.

Klaus-Werner Jörg received the Dipl.-Inform. and the Dr.-Ing. degrees from theDepartment of Computer Science at theUniversity of Kaiserslautern, Germany, in1988 and 1993, respectively. Until early2000, he worked as a researcher andlecturer in the field of autonomousrobots, sensing and machine learn-ing at the University of Kaiserslautern.Since then, he is a member of the

executive board of company Androtec GmbH, Waldfischbach-Burgalben, Germany.

Ewald von Puttkamer, born in 1936,received his diploma degree in physicsfrom the University of Freiburg, Germany,in 1965 and his Ph.D. in physics from thesame university in 1969 with a thesis onphotoion–photoelectron coincidence mea-surements. In 1969, he went to the Uni-versity of Mainz and in 1970 to the newlyfounded University of Kaiserslautern. In1975 he became university professor at the

Computer Science Department. Since the early 1980s his researchinterests are in the field of autonomous mobile robots.

Documents

Reference scan matching for global self-localization