Autonomous Docking for an eROSI Robot Based on a V

AutonomousDockingforaneROSIRobotBasedonaVisionSystemwithPointsClusteringTechnicalReportDepartmentofComputerScienceandEngineeringUniversityofMinnesota4-192EECSBuilding200UnionStreetSEMinneapolis,MN55455-0159USATR07-003AutonomousDockingforaneROSIRobotBasedonaVisionSystemwithPointsClusteringHyeunJeongMin,AndrewDrenner,andNikosPapanikolopoulosJanuary26,2007AutonomousDockingforaneROSIRobotBasedonaVisionSystemwithPointsClusteringHyeunJeongMin,AndrewDrenner,andNikolaosPapanikolopoulosDepartmentofComputerScienceandEngineeringUniversityofMinnesotaJanuary25,2007AbstractThispaperpresentsanautonomousdockingsystembasedonvisualcuesonadockingstation.Autonomousdockingisessentialforlargescaleroboticteamstobedeliveredbylargerrobots,recovered,recharged,andredeployedforcontinuousoperation.Usingacomputervisionbasedapproach,weidentifycuestolineupfordockingbyextractingcornerpixelsandcombiningthisinformationwithcolorinformation.PotentialtargetpointsareextractedandclusteredusingEuclideandistanceintheimageplane.Usingtheseclustersofpointstheappropriatemotionbehaviorisselectedtorepositiontherobotintothedesiredpositionandorientation.ThispaperwillpresentexamplesofthisimplementationusinganeROSIrobotwhichusesonlyvisiontonavigate.1IntroductionThereisagreatdealofinterestinmulti-robotteams,specificallywithapplicationsinmulti-robotcollaboration,targettracking,robotplatooning,orurbansearchandrescue.Onemajoraspectofthisresearchishowtodeliverandrecoverroboticteamstotheareaofinterest.In[1]robotsfollowoneanothertoanareaofinterestwhilein[2]larger“motherrobots”areusedtoshipthesmallerrobotsintoposition.Autonomousdockingisnecessarytomovetherobotstoadesiredorientationandpositioninorderfortherobottoresupplyordeliversamples.Forexample,in[2],therobotsdockinordertorechargetheiron-boardbatteryforadditionalruntime,butthisdockingmayalsobeusedtodeliversoilorairsamplesforanalysiswhichmayrequirelargersensorsthanalltherobotscantransport.Vision-baseddockingisachallengingproblemastherearemanyenvironmentalvariablessuchaslightingandshadowsthatmakeidentificationofthetargetsproblematic.Theseproblemscanbefurthercompoundedonsmallerroboticsystemswhichlacksufficienton-boardprocessingandmusttransmittheirvideoforremoteprocessing.Thistransmissionprocesscanaddadditionalnoiseanddistortiontotheimage.Dealingwiththistypeofnoiseiscriticalinidentifyingthecorrectpositionandorientationfordocking[3].However,evenwiththelimitationsimposedbynoisyimagery,thevisionsystemcanbeapowerfultoolforthesetypesofapplicationsasitprovidesawealthofinformationandisarelativelyinexpensiveformofsensor.Thispaperpresentsatargetdetectionalgorithmbaseduponthecombinationofcornerdetectionandcolorcues.Potentialtargetsarethenclusteredtoincreasethelikelihoodofpropertarget1identification.Utilizingmultipleconsecutiveimages,themethodbecomesrobusttothenoisethatmaybeintroducedthroughtransmissionforremoteprocessing.Figure1showsasampleconfigurationofaneROSIapproachingasetoffeaturesdenotingadockingstation.Onceasetoftargetsisidentified,abehaviorselectionsystemisactivatedtocontrolthemotionoftherobotasitmaneuverstothedockingstation.Astherobotismoving,activeperceptionoftheenvironmentisusedtogenerateadaptivebehaviors[4].Adaptivebehaviorsarenecessaryastheremaybeodometryerrorsduetowheelslippage,terrainfeatures,orimprecisemotorcontrol.Thus,behaviorsarecontinuallymonitoredandmodifiedtoensurethatthesystemisrobusttonoisefromthevisionsystemandpotentiallyunreliablemotioncontrol.Figure1:AneROSIrobotandthetargetfordocking.Theapproachwepresenthasseveralnovelaspectswhichmakeitadvantageousoverotherapproaches:•ComputationalComplexity-Oursystemreducescomputationalcomplexitywhenhandlingimagessinceourimageprocessingonlyextractsaround20candidatepixelsforthetargetpoints.•TargetRepresentation-Thetargetobjectisrepresentedwithpoints.Weestimatethepointclusterstoextractatargetobjectandthemethodprovidestheexactinformationforthecornerswhichwillbeusedfordeterminingtheappropriatemotionbehavior.•TargetShapeIndependence-Theapproachcanutilizetargetsofanyshapesinceweusepointclustering.•CollisionAvoidance-Sincethedesiredtargetsareknownapriori,itispossibletoestimatedepthtothedockingposition.Thisinturncanbeutilizedtoavoidcollisionswiththedockingstationonsuboptimalapproachpaths.Forourexperiments,wewilluseaneROSIrobot,developedattheCenterforDistributedRoboticsattheUniversityofMinnesota[5].TheeROSIisatwo-wheeled,differentiallydrivenminiaturerobotthatisequippedwithaminiaturecamera.However,duetothesmallsizeoftheeROSI,itlacksthecomputationalpowertoprocessvideoon-board.FurtherinformationontheexperimentalsetupwillbediscussedinSection4.Therestofthispaperwillbedividedasfollows.RelatedliteraturewillbediscussedinfurtherdetailinSection2.Section3willdiscusstheframeworkforalignmentfordockingalongwiththespecificsoftheclusteringapproaches,selectionsofbehaviors,andposeestimationfromfeaturesselectedintheimageplane.ExperimentalresultswillbepresentedinSection4.ConclusionsandfuturedirectionsfortheworkwillbediscussedinSection5.2RelatedWorksTheprocessofdockingcanbedividedintotwomainapproaches.The