Mobile Robot Navigation in a Corridor Using Visual Odometry
Bayramoglu, Enis; Andersen, Nils Axel; Poulsen, Niels Kjølstad; Andersen, Jens Christian; Ravn, OlePublished in:
Prooceedings of the 14th International Conference on Advanced Robotics
Publication date:2009
Document Version
Publisher's PDF, also known as Version of recordLink to publication
Citation (APA):
Bayramoglu, E., Andersen, N. A., Poulsen, N. K., Andersen, J. C., & Ravn, O. (2009). Mobile Robot Navigationin a Corridor Using Visual Odometry. In Prooceedings of the 14th International Conference on AdvancedRobotics. (pp. 58). IEEE.
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and 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 you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.
MobileRobotNavigationinaCorridorUsingVisualOdometry
EnisBayramog˘lu∗,NilsAxelAndersen∗,NielsKjølstadPoulsen†,
JensChristianAndersen∗andOleRavn∗
∗Department
ofElectricalEngineering,AutomationandControlGroup,TechnicalUniversityofDenmark,Lyngby,Denmark
Emails:{eba,naa,jca,or}@elektro.dtu.dk
†DepartmentofInformaticsandMathematicalModelling.TechnicalUniversityofDenmark,Lyngby,Denmark
Email:nkp@imm.dtu.dk
robotAbstractgenerationlocalization—Incorporationitsoflocalizationisstudiedofcomputervisionintomobileinformationinthiswork.fromrawItincludestheisfusionwiththeodometricposeestimation.Theimagestechniqueandawithcorridorthenimplementedenvironment.onAasmallmobilerobotoperatingatextraction.animprovedwayofdiscretizationnewsegmentedisusedHoughtransformclassifyThevanishingpointconceptisthenincorporatedforimagelinetoinvolvinglinestoThefindboththeasiterativewellastoestimatetheorientation.Amethodthevanishingeliminationoftheoutliersisemployedodometryfusiondrivenisbetweenachievedthewithvisionpointanextendedbasedandposethecameraposition.Kalmanestimationfilter.AandtheerrorerrormodelisusedfortheodometrywhileadistancesimpleAnappliedextendedmodelwithKalmanconstantnoiseisassumedforthevision.aresystemincluded.toestimateisillustratedTherobustnessodometryfilterasparameters.aparameterestimatorisalsobyperformingandthesimpleprecisionExperimentalnavigationoftheresultstasks.
entireI.INTRODUCTION
Thefieldofrobotvisionisgainingprominenceasthepos-sibilitiesareexplored.Theimportanceofvisioninhumansascomparedtoallothersensespayscredittothat.Manytasksperformedbyhumans,today,requirevisionandtheirautomationcouldbemadepossibleasthecomputervisionfielddevelops.Themostsignificantadvantageofvisionisitscapabilitytoacquireinformationeveninverycomplexenvironments,withoutinterferingwiththesurroundings.Thismakesitaveryflexiblesense.
Ontheotherhand,mobilerobotics,stillinitsinfancy,isoneareawhereflexibilitytowardstheenvironmentismostdesired.Themajorchallengeofmobilerobotics,asthenamesuggests,istonavigatetherobottowhereitneedstobe.Inadditiontotheobviousrequirementofprimarymotioncapabilities,suchasturningandmovingbackwardsandforwards,therobothastosenseitspossiblydynamicenvironment,determineitsownlocationandgenerateamotionplanaccordingly.
Theprojectdescribedinthisarticleaimstocontributetobothfieldsbyapplyingcomputervisiontoperformoneofthemostimportanttasksofmobilerobotics,localization.Thescopeofthisprojectincludesthegenerationofthisinformationfromtheimagesaswellasitsfusionwith
wheelencoders.Asmallmobilerobotoperatinginacorridorenvironmentisthenequippedwiththisposeestimatortoperformsimplenavigationtasksasanindicatorofoverallperformance.
Thefieldsofcomputervisionandmobilerobotlocal-izationhavebeenstudiedextensivelytodate.TheworkofKleeman[8]isagoodexampleofmobilerobotlocalizationwithmultiplesensors,namelyodometryandadvancedsonar.Thecameraposeestimation,independentfromroboticsisalsoaveryactiveresearchfield.Yuetal.[16]usedthetrifocaltensorwithpointfeaturestoestimatethepathofthecamerafromasequenceofimagesandMakadia[11]investigatedthecameraposeestimationrestrictedonaplane.Inthesubjectofmobilerobotlocalizationwithvision,Andersenetal.[2]employedmonocularvisiontoassistthelaserscanner.Lin[9]usedstereovisionandMunguiaandGrau[12]studiedmonocularvisiondirectly.
Previousworkinvestigatingproblemssimilartothisprojectshouldalsobenoted;Tada[15]usesmonocularvisioninacorridorandincorporatesthevanishingpointtofollowthecenterwhileShietal.[14]studiesnavigationinacorridorusinglinesbuttheyaremainlyinterestedinasaferegiontotravelinsteadoftheposeandfollowaverydifferentstrategy.Guerraetal.[6]obtainincrementalmeasurementsfromlinesinenvironmentsunknownaprioriandtheyalsocarryoutexperimentsinacorridor.
II.SOLUTIONOUTLINE
Intheassumedsetup,themobilerobothasasinglecameramountedonitwithoutanyabilitytoturnormovew.r.ttherobotitself.Theactivewheelsoftherobotalsohaveencodersavailablefordeadreckoning.Therobotlocalizationisperformedatacorridor,wherethevisionisusedtoestimatetherobotorientationanditspositionacrossthewidthofthecorridor.Deadreckoning,ontheotherhand,isusedtokeeptrackofboththeorientationandthepositionacrossthewidthandthedepthofthecorridor.Thecorridordimensionsareassumedtobeknownapriori.
Thechoiceofthecorridorastheworkingenvironmenthastwoimportantreasons.First,thecorridorisacommonpartofmostdomesticenvironmentsandbeingabletonavigatein
ithaspotentialonitsown.Second,ithasaveryregularandsimplestructure,makingiteasiertostartthedevelopmentofamoregeneralvisionbasedsolution.
Deadreckoningisalwaysappliedtokeeptrackoftheposeestimationwithagrowingerror.Therefore,foreachrawimagetaken,apriorestimateisavailabletoaidthevisualestimation.Initially,therobotiseitherstartedfromaknownlocationorahighuncertaintyintheposeisassumed.
Visualestimationisperformedusingimagelinesasfea-tures.LinesareextractedusingaformofsegmentedHoughtransform.Theselinesarethenclassifiedw.r.tdirectionusinginvariantenvironmentalinformationandthepriorestimate.Thevanishingpointcorrespondingtothedirectionalongthecorridorisalsocalculatedrobustly.Thelinesarethenmatchedtoeachofthefourlineslyingalongthecornersofthecorridors.Finally,thevanishingpointisusedtoestimatetheorientationwhilethelinematchesareusedtoestimatethetranslation.
WhenthevisualestimateisavailableitischeckedforconsistencyusingthepriorestimateandBayesianhypothesistesting.Ifitpassesthecheck,itisfusedwiththedeadreckoningusinganextendedKalmanfilter(EKF).ThemodelforthedeadreckoningismodifiedtoallowfortheestimationofitsparametersalongwiththeposeitselfresultinginanextendedKalmanfilterasaparameterestimator(EKFPE).Theprocessingofeachimagespansafewsamplingperi-odsofthewheelencoders.Duringthistime,deadreckoningiscontinuedandalltheencoderoutputisrecordedatthesametime.Whenthevisualestimateisavailable,itisusedtorefinetheestimateatthetimeofthetakingoftheimageandtheestimateforthecurrenttimeisrecalculatedusingtherecordedwheelencoderoutput.
Thenextsectiondescribesthevisionbasedpartoftheposeestimationincludinglineextraction,linematchingandposeestimationsteps.SectionIVisconcernedwiththefusionofthevisionbasedestimationwiththeodometry.SectionVpresentstheresultsofthework.
III.VISION
A.LineExtraction
Anedgeimageisfirstcreatedtobeusedinthelineextraction.Cannyedgedetectionalgorithmisusedforthispurpose.Thealgorithmisfirstproposedandexplainedin[4].Itisonlyslightlymodifiedbyperformingthenon-maximaledgepixelsuppressionbycomparisonwithverticalandhorizontalneighborsonly,thatis,excludingthediagonalneighbors.Thismodificationresultsinthinnerlines,fasterandmorepreciseforlineextraction.
SegmentedHoughtransformisthenappliedtotheresult-ingedgeimage.Thereadercouldreferto[1]forHoughtransformusedforgeneralfeatureextraction.ThestandardHoughtransform(SHT)usedforlineextractionisexplainedin[5].[7]givesanoverviewofthevariationsofSHT.SegmentedHoughtransformispreferredinthisworkduetoitshigherspeedandrobustness.
Theexactprocedureistofirstsegmenttheedgeimageinto10x10subimages.Houghtransformisthenapplied
tothoseimages.ThissegmentationcanbeshowntospeeduptheoverallHoughtransformproportionaltothesquarerootofthenumberofsegments.Theperformanceisfurtherincreasedbytheextensiveuseoftablelookupsmadepossiblebythesmallsizeofeachsubimage.Thelinesobtainedfromeachsubimagearethentracedacrossthesubimages.
Thealgorithmwithalltheperformanceoptimizationsresultedin15-60timesspeedincreasecomparedtotheOpenCV1implementationofSHT.TheHoughtransformisalsomodifiedtokeeptrackofsupportingpixelsforeachlinethroughtheuseoflook-uptables.Twolinesegmentsarecombinediftheircombinedsupportingpixelsdescribealinepreciselyenoughwithathreshold.Aswellasbeingarobustcriterionforlinecombination,thisprovidestheendpointsoflinesrobustlyandwithoutanyneedforfurtherprocessing.B.VanishingPointDetection
Ifanumberoflinesareparallelinthe3Dscene,theirprojectionontheimageallmeetatasinglepoint.Thispointisthesocalled”vanishingpoint”specifictothedirectionofthoselines.Thevanishingpointisausefultoolbothforthedetectionofthe3Ddirectionsofimageandforthecalculationofthecameraorientationw.r.tthatdirection.Thevanishingpointisexpectedtositonapointwheremanylinesintersectwithallothers,ifthereareenoughsupportinglines.Whenalltheintersectionpointsbetweentheimagelinesarecalculated,adenseclusterissupposedtobeformedaroundthevanishingpoint.Usingtheavailablepriorestimate,itisalsopossibletocalculateanestimateforthevanishingpointalongwithanuncertainty.
Givenadirectionintheworldcoordinatesdescribedbytheunitvectorv,thisvectorisfirsttransformedtothe
imagecoordinateswhereitresultsinvi.Notethatv
iwillbeafunctionofthecameraorientationbutnotthecameratranslation.Thecoordinatesofthevanishingpointontheimagewillthenbegivenby:
ixv=vix
vyvi(1)
z,yv=vi
z
Inthisimplementation,firsttheintersectionpointsare
calculated.Theyarethenfilteredusingthevanishingpointestimateobtainedfrom(1).Theactualvanishingpointiscalculatedbyiterativelyremovingthefurthestpointfromthecenterofgravityoftheremainingpoints.Whenthenumberofpointsisreducedtoacertainthreshold,thecenteroftheremainingpointsisusedasthevanishingpoint.Equation(1)isthenusedtocalculatethecameraorientation.Thisschemehasproventobeveryrobust.Foranalternativemethodofvanishingpointdetection,thereadershouldreferto[3].Itisimportanttonotethat(1)providestwoconstraintswhereasthecompletecameraorientationisdescribedbythreeparameters.Inthecaseofthemobilerobot,thecameraisconstrainedtoturnaroundasingleaxis,thereforethosetwoconstraintssuffice.
1An
opencomputervisionlibraryforC/C++,originallydevelopedby
Intel.
C.PoseEstimation
Asthevanishingpointcontainsinformationaboutonlytheorientationofthecamera,imagelinesarealsomatchedtotheknown3Dlinestoobtainconstraintsaboutthecameratranslation.Eachlinecontributesoneconstraintforthepositionalongthewidthofthecorridor.
AssumethatanimagelineisdescribedbytheCartesianlineequationgivenin2.
ax+by=c
(2)
Wherea,bandcarecalculatedduringlineextraction.Thenifthepointpisapointonthelinein3Dandthevectorvisthedirectionoftheline,thetransformationof
thosevectorstotheimagecoordinates(pi,v
i)satisfiesthefollowingconstraints:
avi++bviapx
ibpy=cvixizy=cpi
(3)
z
Ideally,thefirstofthoseissimplysatisfiedbycalculatingtheorientationusingtheestimatedvanishingpoint.Thesecondconstraintisthenenoughtosolveforacandidatevaluefortheposition.
Inthiswork,thelinesalongthecorridorcornersareused.Thelinespassingthroughthevanishingpointareclassifiedtobealongthecorridor.Inordertomatchthoselinestoaparticularcorner,imagelinesareinvestigatedfortheirpositionwithrespecttothevanishingpoint.Therobotisknowntobeinsidethecorridor,thereforethoselineslyingbelowthevanishingpointareknowntobeonthefloorandthoselineslyingtoleftofthevanishingpointareknowntobeontheleftwallandviceversa,assumingthatthelinesactuallycorrespondtoanactualcorner.
Thepositionconstraintforeachlineaftermatchingissolvedtoobtainapositioncandidate.Thesameiterativeeliminationofthefurthestelementisagainusedtoarriveataconsistentsetofpositioncandidates,andtheresultisusedasthevisualestimatealongwiththepreviouslycalculatedorientation.
D.ConsistencyCheck
Underrarecircumstancesthevanishingpointisdetectedoveraspuriouscandidatecluster,orbycoincidenceanin-correctsetoflinesalongthecorridorconstitutesaconsistentsetofcorners.Inthesecases,theerrorofthevisualestimateismuchhigherthanwhatisexpectedofaregularestimate.SuchahigherrorcausesabiasattheEKFoutputforalongtimeduetothelowassumeduncertainty.Insteadofsimplyincreasingtheuncertaintyinthevisionerrormodel,atwohypotheseserrorcheckingisusedtorejectsuchfaultyestimates.
Thefirstofthosetwohypothesesisthatthemeasurementisaregular,accurateone.Thishypothesishasthesameerrordistributionastheerrormodelusedinfusion.Thesecondhypothesisisthatthemeasurementisafaultymeasurement.Inthiscasetheerrorisassumedtobemuchlarger.Thesec-ondhypothesisisgivenalowerpriorprobabilitycomparedtothefirst,sincesuchmeasurementsoccurrarely.
Theprobabilityofthefirsthypothesisbeingcorrect,given
theactualmeasurementandthepriorestimate,alongwithitsuncertaintyiscalculated.Ifthisprobabilityisabove95%theestimateisacceptedandusedforfusionasdescribednext.
IV.SENSORFUSION
ExtendedKalmanfilterischosenforthefusionofthetwosourcesofinformationavailableforlocalization.Thischoicerequiresanerrormodeltobedefinedforbothsources.A.ErrorModels
Asimpleexperimentallytunedandvalidatederrormodelisdefinedforthevisualinformation.Accordingtothismodel,theorientationismeasuredwithanindependentaddi-tiveGaussianerrorhavingzeromeanand0.01radstandarddeviation.ThetranslationmeasurementisalsoassumedtohaveanindependentadditiveGaussianerrorwithzeromeanand3cmstandarddeviation.
Theerrormodelchosenforthedeadreckoningisde-scribedin[8].Thismodelcouldbesummarizedasfollows;thedeadreckoningequationisgivenin4.Inthisequationθistheorientationandx,yarethepositionestimates.llrandlaretheleftandrightwheeltraveleddistanceswhileBistheeffectiveseparationbetweenthewheels.⎡
⎣θlθ(k)+
lr(k)−ll(k)
⎤
x(k(k+⎤⎡
+1)
1)⎦=⎢r(k)+ll(k)
B
lr(k)y(k+1)⎣x(k)+y(k)+l(k)+lcos(θ(k)+
r2
2−ll(k))⎥⎦2
l(k)
sin(θ(k)+
lr(k)2−B
B
ll(k)
)(4)
Thesourcesoferrorinthisestimateupdateequationareassumedtobeduetouncertaintyonluncertaintiesarefurtherassumedtobeindependentl,lrandB.TheGaussianwithzeromeanandvariancesproportionaltothetraveleddistance.Thismodelensuresthattheresultinguncertaintyforapathisindependentofthenumberofsamplestakenduringit.
B.FusionUsingEKF
TheKalmanfilterisanoptimalstateestimatorforafinitelinearsystemwiththeinitialstateestimatesanderrorsourcesjointlyGaussianinnature.ExtendedKalmanfilterisanextensionofthisfiltertonon-linearsystemsbylinearizingthestatespacedescriptionlocally.Anindepthanalysisofbothcouldbefoundin[13].
ThedeadreckoningisusedintheKalmanfilterinplaceofthesystemstatetransitionequation,althoughitisalsoameasurement.Thisapproachisfollowedbymanyauthorsanditcanbevalidatedmathematically.Thevisualestimationentersthefilterasadirectmeasurementoftheorientationandtheycoordinate.
ItisimportanttonotethattheresultingfilterisslightlydifferentfromaconventionalEKFasthemeasurementisnotappliedateverystatetransitionsample.Instead,itisappliedwheneveravisualestimateisavailable.
C.ModifyingtheSystemforEKFPE
AnextendedKalmanfilterasaparameterestimator(EKFPE),asthenamesuggests,isamodificationoftheEKFsothatitestimatessomeofthesystemparametersalongwiththestates.Thisisachievedbyaugmentingthestatevectorwiththeparameterstobeestimated.Thestatetransitionequationisalsoaugmentedwiththesenewstatesthatdonotchangeoversamples.Aninitialuncertaintyisalsoassumedoverthesestatestoallowforestimation.EKFPEisanalyzedindetailin[10].
Theodometryequationin(4)usedasthestatetransitionequationismodifiedasfollows:⎡
θ(k)+
lr(k)−ll(k)
⎤
⎢θ(k+1)
⎤⎡
⎢lB(k)
r(k)+ll(k)
⎢xcos(θ(k)+lr(k⎢⎢y((kk++1)⎥1)⎥⎢⎥⎢⎢x(k)+l(k)+2
rk⎥⎢y(k)+r(k+1)2
ll(k)
sin(θ(k)+
l2B)−ll(k)⎥k2B)−(k(l))⎥r(kkl)
(k)⎢)⎥⎥⎥⎣⎥=⎢k(k+1)⎥⎦⎢l⎢r(k)⎥B(k+1)⎣kl(k)⎥⎦
B(k)
(5)
Here,thenewvariableskfortheleftrandkandrightlarethetraveleddistanceperencodertickwheels.Averysmallsourceoferrorforthesestatesinthestatetransitioncouldalsobeaddedtoallowforslowlyvaryingparametersinthelongrun.
V.RESULTS
A.VisualEstimation
ThestepsofvisualposeestimationareshowninFig.1.Fig.1ashowstherawimagetakenbythemountedcamera.TheedgemapobtainedusingCannyedgedetectorisdisplayedinFig.1b.TheresultofthelineextractionisshowninFig.1c,wheretheextractedlinesaredrawninred.Fig.1dillustratesthecalculationofthevanishingpoint.Thetinyblue’x’marksarewheretheextractedlinesintersect.Atthecenterofthatimageadenseclusterofthosemarksarevisible.Thelargergreen’x’markatthecenterofthatclusteristhedetectedvanishingpoint.
Fig.1edisplaystheendresultoftheprocessing.Thecyanlinesaretheonesclassifiedtobealongthecorridor.
Thevisualestimateforthiscaseisfoundtobe20.8cmtotherightofthecorridorcenter.Theactualpositionismeasuredwitharulerandfoundtobe22cminthesamedirection.
Theerrorintheorientationismuchhardertomeasuresinceitisknowntobeverysmallaslongasawrongclusterisnotfound.Fortunately,thisrarelyhappensduetothepreprocessingappliedbeforevanishingpointdetection.Whenithappens,theestimatefailstheconsistencycheckwiththepriorestimateanditisnotused.Inaregularimagewherethevanishingpointisdetectedinsidetherightcluster,thedeviationcouldbeassumedtobebelow2pixels,whichtranslatestoroughly0.3degrees.
TherobotplatformconsistsofaCPUrunningwith500MHzto1200MHzclockspeeddependingontheage
(a)Originalimage(b)Edgeimage
(c)LineExtraction(d)VanishingPointDetection
(e)CompletelyAnalyzedImage
Fig.1:Visualizationofthestepstakenforvisualposeestimation
oftherobot.Theentirevisualestimationtakesfrom50msto100msforcompletion,dependingontheparticularrobotandtheimage.
AfewotherexamplecasesareshowninFig.2.Theyillustratedifferentcasespossiblyencounteredduringopera-tion.ThevisualestimateworkswellinFigs.2a,2band2cwithpositionerrorsof2.5cm,0.2cmand0.9cmrespectively,despitedifferentchallenges.Fig.2disarareexample,wherealthoughtheimageiswellsuitedforestimationandthevanishingpointiscorrectlyfound,theerrorisashighas11error.8cm.modelTheseforrarethevisualcasesareestimateactuallyassumesthereason3cmwhystandardthedeviationforthepositionerroralthoughtheerrorislowerusually.
B.SimpleNavigationTasks
Theperformanceoftheentiresystemisevaluatedbyperformingsimplenavigationtasksusingthelocalizationdescribedinthisarticle.Thefirsttaskistodrivebackandforth10mfortwocompleteturns.Fig.3containsrelevantinformationforthistask.Fig.3aplotsthevisualestimates(greendots),theoverallestimate(redline)andtheactual
(a)Partiallyoccluded(b)Toofewlines
(c)Lookingaway(d)Acasewithhighpositionerror
Fig.2:Variouscasesforvisualposeestimation;thesearein-tendedtobeinterestingcasesencounteredduringoperation.Theestimationerrorsforimagesa,bandcareallbelow3cmdespitevariousdifficulties.Ontheotherhand,theerrorforimagedis11.8cmalthoughtherearenodisruptingfactors,arareoccurrence.
rulermeasurements(blueline)together.Fig.3bdisplaysthedeadreckoningaloneforthesametask.Notehowtheactualerrorstayswithin±3cmwhiledeadreckoningalonequicklylosestrackofthepose.
Fig.4displaystheresultsoftwomoretasks.ThetaskgiveninFig.4aistomoveonazigzagshapedpath.ThenexttaskgiveninFig.4bistodrivestraight,butduringtheexecutionofthistaskthecameraisblindedalongthepathsegmentbetween6mand13m.Theblackdotsinthefigureillustratethesampleswherevisualestimatesarediscardedautomatically.Notehowtherobotdivergesfromitspathwhileitisblindedandalsohowitquicklyincorporatestheimagesoncethevisionisback(watchtheredline).
Finally,itisimportanttoobservetheevolutionofthedeadreckoningparameters.Fig.5isrecordedduringalongzig-zagtypenavigationtask.Thetaskspans2500visualestimatesandthedeadreckoningparametersareestimatedateachsample.ThefigureshowstheplotofestimatedB,theeffectivewheelseparationandtheestimatedR,whichistheratioofthetraveleddistancesperencodertickfortheleftandrightwheels.Inthisexperimenttheinitialvaluesaredeliberatelysettowrongvalues(Bto0.22cminsteadof0very.27cmquickly,Rtowhile1.1insteadBtakesof0longer.
.99).ObservehowRconvergesVI.CONCLUSION
Avisualodometrymethodformobilerobotlocalizationispresentedhere.Themethodreliesonstraightlinesalongthecorridor,whosewidthandheightareassumedtobeknownapriori.Imagelinesarethenrobustlymatchedtothesetoobtainposeconstraints.TheposeinformationisthenfusedwithdeadreckoningusinganextendedKalmanfilter.
Thepotentialbehindrobotvisionisgenerallyacknowl-edgedduetoboththelowcostofcommerciallyavailable
0.08combinedvisual0.06ruler0.040.02)m0(noitis−0.02op y−0.04−0.06−0.08−0.1−0.12 012345678x position(m)(a)Positiondatafromvarioussourcesforthefirstexperiment4)m2(noitis0op y−2−40510x position(m)(b)Thepathrecordedbyodometryonly
Fig.3:Localizationdatafordrivingstraight.Notehowtheodometryaccumulatesmorethan1merrorforapathsegmentshorterthan10m.Thefusedestimatemaintainslessthan3cmerrorindependentofdistancetravelled.
cameras,theflexibilityofusingvisionandthehighinfor-mationcontentinimages.Thisworkinvolvesthesuccessfulapplicationofvisionforrobustandrealtimeestimationofposeonamoderatesystem.Furthermore,themeasuredprocessingtimeallowstheuseofslowersystemsrunningwithclockspeedsdowntoroughly200MHz.
Thelineextractionstepisthebackboneofthemethodfollowedinthiswork,becausethelinesaretheprimarysourcesofinformation.Thisstepisalsothemosttimecon-sumingone.Thereforeanewfastlineextractionalgorithmisdevelopedwhichisdemonstratedtoberobustindetectingeventheshorterlines.
Thevanishingpointdetectionandlinematchingpartsbothemploytheiterativefurthestpointremoval.Thisensuresthattheyarehighlytoleranttooutliers.Furthermore,theiterativefurthestpointremovalisimplementedusingaquadtreedatastructuresothattheprocessingtimescaleswellwiththenumberofvanishingpointcandidates.
Visualestimationanddeadreckoningbothrelyontheresultsofeachother.However,forrobustness,bothofthemaredesignedtobetoleranttomoreerrorthantheotheronegenerates.Anoveralldemonstrationofthisisprovidedintheresultssectionwheretherobotisblindedforaslongasa6mpathsegment.Duringthissegmentitmanagestostayonitstrackwithinanacceptableerrorbound.Whentheblindingisremovedittakeslessthanhalfametertorelocateitself.On
0.3 combinedruler0.2visual0.1)m(noitis0op y−0.1−0.2 0510152025x position(m)(a)Positiondataforazig-zagshapeddrive0.15 combineddiscarded0.1visualruler0.05)m(n0oitisop −0.05y−0.1−0.15−0.20 51015202530x position(m)(b)Positiondataforthepartiallyblindeddrive
Fig.4:Additionalnavigationtasks
theotherhand,visionisalsodemonstratedtosuccessfullycompensatefortheerrorsinthedeadreckoningparametersaswellasestimatingthem.
Theaccuracyofthesystemisalsodemonstratedthroughexperiments.Themeasuredpositionerrorismaintainedbelow3cmandtheorientationerrorisestimatedtobebelow1degree.
VII.FUTUREWORK
Thedevelopedmethodpresentsmanyopportunitiesinconjunctionwithothersystems.OthersensorscouldbesmoothlyintegratedthroughtheExtendedKalmanFilter.Themethoditselfcouldalsobeinsertedinothervisionsystemswherevanishingpointsareoccasionallyavailable.Thiswouldprovidethehostsystemwithanindependentsourceofmeasurement.Althoughitisdevelopedforastraightcorridor,itcouldbeusedinmorecomplexcorridorenvironmentsprovidedthatthecornersandjunctionsarehandledwithauxiliaryalgorithms.
Ongoingresearchaimstoextendtheproposedsolutiontoincorporateagenericsetoflinesthatarenotrestrictedtobeparallel.Theuseofmultiplecamerasinconjunctionisalsoinvestigated.Themotivationbehindtheextensionistoenablelocalizationbasedonthismethodinlessregular
0.280.26X: 2500)Y: 0.2709m(B0.240.2205001000150020002500Sample number1.1R1.05X: 25001Y: 0.990505001000150020002500Sample numberFig.5:Theevolutionoftheodometryparametersenvironmentswithavailablelines.
REFERENCES
[1]A.S.Aguado,M.E.Montiel,andM.S.Nixon.Arbitraryshapehough
transformbyinvariantgeometricfeatures.IEEEInternationalConfer-enceonSystems,Man,andCybernetics.ComputationalCyberneticsandSimulation,3:2661–2665,1997.
[2]JensChristianAndersen,NilsAxelAndersen,andOleRavn.Vision
assistedlaserscannernavigationforautonomousrobots.InProceed-ingsofthe10thInternationalSymposiumonExperimentalRobotics2006(ISER’06);39,pages111–120.Springer-VerlagBerlin,2008.Presentedat:10thInternationalSymposiumonExperimentalRobotics2006(ISER’06):ISER’06,2006.
[3]S.T.Barnard.Interpretingperspectiveimages.ArtificialIntelligence,
21:435–62,1983.
[4]J.Canny.Acomputationalapproachtoedgedetection.IEEE
TransactionsonPatternAnalysisandMachineIntelligence,8:679–98,1986.
[5]RichardO.DudaandPeterE.Hart.Useofthehoughtransformation
todetectlinesandcurvesinpictures.Commun.ACM,15(1):11–15,1972.
[6]JJGuerreroandCSagues.Uncalibratedvisionbasedonlinesfor
robotnavigation.MECHATRONICS,11(6):759–777,SEP2001.
[7]J.IllingworthandJ.Kittler.Asurveyofthehoughtransform.
ComputerVision,Graphics,andImageProcessing,44:87–116,1988.[8]L.Kleeman.Advancedsonarandodometryerrormodelingfor
simultaneouslocalisationandmapbuilding.IEEE/RSJInternationalConferenceonIntelligentRobotsandSystems,1:699–704,2003.[9]Huei-YungLinandJen-HungLin.Avisualpositioningsystemfor
vehicleormobilerobotnavigation.IEICETransactionsonInformationandSystems,E89-D:2109–16,2006.
[10]L.Ljung.Asymptoticbehavioroftheextendedkalmanfilteras
aparameterestimatorforlinearsystems.IEEETransactionsonAutomaticControl,24:35–50,1979.
[11]AmeeshMakadia,DinkarGupta,andKostasDaniilidis.Planarego-motionwithoutcorrespondences.Proceedings-IEEEWorkshoponMotionandVideoComputing,MOTION2005,pages160–165,2007.[12]R.MunguiaandA.Grau.Monocularslamforvisualodometry.IEEE
InternationalSymposiumonIntelligentSignalProcessing,pages1–6,2007.
[13]MariaIsabelRibeiro.Kalmanandextendedkalmanfilters,2004.
[14]omni.isr.ist.utl.pt/WenxiaShiandJ.Samarabandu.∼mir/pub/kalman.pdfCorridorlinedetection.
forvision
basedindoorrobotnavigation.CanadianConferenceonElectricalandComputerEngineering,pages1988–1991,2006.
[15]N.Tada,T.Saitoh,andR.Konishi.Mobilerobotnavigationbycenter
followingusingmonocularvision.SICEAnnualConference,pages331–335,2007.
[16]YingKinYu,KinHongWong,M.M.Y.Chang,andSiuHangOr.
Recursivecamera-motionestimationwiththetrifocaltensor.IEEETransactionsonSystems,Man,andCyberneticsPartB:Cybernetic,36,2006.
因篇幅问题不能全部显示，请点此查看更多更全内容