外文文献翻译基于激光测距仪的行人跟踪.docx
《外文文献翻译基于激光测距仪的行人跟踪.docx》由会员分享,可在线阅读,更多相关《外文文献翻译基于激光测距仪的行人跟踪.docx(9页珍藏版)》请在冰豆网上搜索。
外文文献翻译基于激光测距仪的行人跟踪
ActivePedestrianFollowingUsingLaserRangeFinder
I.INTRODUCTION
Theabilityofrobotstotrackandfollowmovingtargetsisessentialtomanyreallifeapplicationssuchasmuseumguidance,officeorlibraryassistance.Ontopofbeingabletotrackthepedestrians,oneaspectofhuman-robotinteractionisrobot’sabilitytofollowapedestriantargetinanindoorenvironment.Therearevariousscenarioswheretherobotcanbegiveninstructionssuchasholdingbooksinalibraryorcarryinggroceriesatastorewhilefollowingthepedestriantarget.
ThekeycomponentsofmovingtargetfollowingtechniqueincludeSimultaneousLocalizationandMapping(SLAM),DetectingandTrackingMovingObjects(DATMO),andmotionplanning.Moreoftenthannot,therobotsarerequiredtooperateindynamicenvironmentswheretherearemultiplepedestriansandobstaclesinthesurroundings.Consequently,trackingandfollowingaspecifictargetpedestrianbecomemuchmorechallenging.Inotherwords,thefollowingbehaviorsmustberobustenoughtodealwithconstantocclusionsandobstacleavoidances.
Whendesigningthefollowingalgorithm,oneintuitiveapproachistosetthetargetlocationasthedestinationfortherobot.However,thisapproachcaneasilyleadtolosingthetargetbecauseitdoesnotreacttothetarget’smotionnorconsiderthevisibilityproblem(sincethetargetmaybeoccludedbyobstaclesandbecomeinvisible).Forachievingrobusttargetfollowingandtracking,therobotshouldhavetheintelligenttopredicttargetmotionandgatherobservationsactively.
Inthispaper,weproposeamovingtargetfollowingplannerwhichisabletomanageobstacleavoidanceandtargetvisibilityproblems.Experimentalresultsareshowntocomparetheintuitiveapproachwithourapproachandprovetheimportanceofactiveinformationgatheringinplanning.Thispaperisorganizedasfollow:
SectionIIdiscussesrelatedworksofDATMOandplanningalgorithms.SectionIIIdescribesourDATMOsystemandintroducesourtargetfollowingplanner.Lastly,SectionIVillustratestheexperimentalresults.
II.RELATEDWORKS
TherearevariousapproachestodetectandtrackmovingobjectssuchasbuildingstaticanddynamicoccupancygridmapsproposedbyWolf&Sukhatme[1],findinglocalminimainthelaserscanasinHoriuchietal.’swork[2]orusingmachinelearningmethodsinSpinelloetal.’swork[3].MostofDATMOapproachesassumethattherobotisstationaryorhasperfectodometry.Whentrackingmovingobjectsusingmobilerobots,ithasbeenproveninWangetal.’swork[4],thatSLAMandDATMOcanbedonesimultaneouslyifthemeasurementscanbedividedintostatic
anddynamicparts.Inourwork,weimplementaDATMOalgorithmwhichissimilartotheoneinMontesanoetal.’swork[5].Ascanmatchingmethodisusedtocorrectrobotodometryandmovingpointsaredetectedbymaintainingalocaloccupancygridmap.Moreover,ExtendedKalmanFilter(EKF)isappliedtotrackthemovingobjects.
Thispaperaimstosolvemovingtargetfollowingproblemwiththeexistenceofobstacles.Fornavigatinginstaticenvironments,therearemanysuccessfulworkssuchasFoxetal.[6],Ulrich&Borenstein[7]Minguez&Montano[8],Seder&Petrovi’c[9].However,thosemethodsaredesignedtoreachafixedgoalandassumethattheenvironmentandrobotstatesarefully-observable.Applyingtraditionalobstacleavoidancealgorithmsonthetargetfollowingtaskcanfaileasilybecauseamovingtargetcanchangeitsspeedandmovingdirectionatanytimeandthetargetcanbeoccludedbyobstacles.Forplanningunderimperfectperception,Partially-ObservableMarkovDecisionProcess(POMDP)providesageneralframework.However,usingPOMDPtocomputeoptimalpoliciesareusuallyverycomputationallyexpensivebecauseithastocomputeaplanoverbeliefspace(typicallyN-1dimensionalforanN-stateproblem.).ForapplyingPOMDPsonpracticalproblems,mostworksaimedonreducingthedimensionofbeliefspace.SuchlikePBVIPineauetal.[10],AMDPRoyetal.[11]andMOMDPSylvieetal.[12].ThosemethodsaremuchfasterthanoriginalPOMDP,buttheircomputationalcomplexityarestilltoohightodoreal-timeplanning.Ourmethodinthispaperismorelikeasub-optimalbutfastapproach:
assumingthattherobotalwaysreceivesthemostpossibleobservationsandplanapathwhichcanreachthegoalandminimizetheuncertaintyonparticularstates.Forexample,inPrentice&Roy[13],therobotaimsonreachingthegoalwithminimumuncertaintyonitsposition,soitmaychooseapathwhichislongbuthasenoughlocalizationlandmarks.
Inthispaper,weproposeamotionplannerformovingtargetfollowing.TheplannerusesanextensionofdynamicwindowapproachpropsedbyChou&Lian[14]tofindcollision-freevelocitiesandchooseapropervelocityusingheuristicsearch.Costfunctionsaredesignedforminimizingthedistancebetweenrobotandtargetandmaximizethepossibilitythattherobotcankeepobservingthetargetinafixedtimehorizon.Additionally,weapplytheconceptinnearnessdiagramalgorithmsuchastheoneinMinguez&Montanos’work[8]forcomputingabetterestimationofthedistancebetweenrobotandtargetandthereforeachieveasmooth,non-hesitatingperformance.
III.MOVINGTARGETFOLLOWINGANDOBSTACLEAVOIDANCE
ItisessentialthatourDATMOsystemisabletotrackthemovingtarget,pedestrianinthiscase,withgreataccuracy.Themoreprecisepedestrianlocationacquired,thebettertherobotperformswhenfollowingthetarget.
A.DetectingandTrackingMovingObjects
Fordetectingmovingpointsinthelaserdata,weadopttheconceptofoccupancygridmapproposedbyWolf&Sukhatme[1].Alocaloccupancygridmapisaintainedandusedtodifferentiatethemovingpoints.Forrobotlocalization,ascanmatchingtechniquecalledIteratedClosestPoint(ICP)isusedtoacquirerobotpose.Themovingpointsarefilteredoutofthedatapriortoscanmatchinginordertomaintaintheposeaccuracy.Thedetectedmovingpointsaresegmentedintonumerousclustersanddeterminediftheybelongtoapedestrianusingfeaturessuchasmotionvelocity,localminima,andsize.Finally,eachofthepedestrianistrackedusingExtendedKalmanFilter(EKF)whichsolvestheocclusionproblemandprovidesusthecomputationaladvantageinrealtimeperformance.
B.FollowingMethod
Ourgoalistoaccomplishmovingtargetfollowingwiththeexistenceofobstacles.Areasonablesolutionistoseeitasapathplanningproblemanduseobstacleavoidancealgorithmstofindcollision-freeactions.Inthispaper,weapplyourpreviouswork,DWA*,Chouetal.[10]astheobstacleavoidancealgorithm.
TheprocedureofDWA*isshowninFig.1(a),itisatrajectory-rolloutalgorithm.TherightsideofFig.1(a)showstheprocedureforcomputingpropermotioncommands.First,theenvironmentinformationisrealizedasintervalconfigurationforfasterprocessing.Eachintervalvaluerepresentsthemaximumdistancecanberaveledbytherobotonacertaincirculartrajectory.Second,theintervalsareclusteredasnavigableareas.Third,foreacharea,acandidatevelocityisdeterminedaccordingtoanobjectivefunction.Foreachcandidatevelocity,anewrobotpositioniscomputedasanewnodeandsavedinatrajectorytree.Thenanodewiththeminimumestimatedcostvaluewillbeextractedasthebasenodeforgeneratingnewnodes.Theprocedureisrepeateduntilgoallocationisexpandedorthetreedepthreachesacertainvalue.Afterthetreeexpansionstops,thedeepestnodeisdeterminedasatemporalgoal,andthepresentcandidatevelocitywhichcanleadtherobottothetemporalgoalisselected.
Inourwork,weusedtwodifferentmethodsforintegratingDATMOandDWA*:
pseudogoalmethodandtrajectoryoptimizationmethod.Pseudo-goalmethodisamoreintuitiveapproachwhichdoesnotconsiderthetargetvelocitywhenfollowing.Anotherapproachistrajectory-optimizationwhereatrajectory-rolloutcontrollerisusedtoapproximateapredictedtargettrajectoryandmaximizetargetvisibility.Theirperformanceswillbeshownandcomparedinthenextsection.
1)Pseudo-goalmethod
Whenimplementingpeoplefollowingalgorithm,anintuitiveapproachissettingthepresentlocationofthemovingtargetasthegoalofnavigationalgorithm.Considerthetrackedtargetatananglewithrespecttotherobot,DWA*algorithmwillgenerateanangularandtranslationalvelocitywhichproducesanarc-liketrajectory.Ifthegoaliswithinacloseproximityoftherobot,theangularvelocitywillbesmallandtherc-likerouteoftenresultsinanindirectordetourpathfortherobottoreachthegoal(asshowninFig.2).Onthecontrary,ifthegoalissetfurtherawayfromtherobotatthesameangle,anymovementofthegoalwillresultinamuchlargerdisplacementchangecomparetowhenthegoalisclosetotherobot.Therefore,theangularvelocitywouldincreaseinresponsetothesubstantialchangeofthegoallocation.Consequently,thetrajectorybecomesmoredirecttowardthegoal.
Inthepseudo-goalmethod,thespaceinfrontoftherobotisdividedinto7trajectoriesat35_,55_,70_,90_,110_,130_,145_asshowninFig.3.Apseudogoalisset3timestheoriginaldistancebetweenthegoalandrobotalongeachtrajectory.Afteracquiringthepedestrianlocation(redcircleinFig.3),thepseudogoalwithtrajectoryclosesttothepedestrianlocationisthenselected.Bysettingthegoalfurtherawayfromtherobot,itremediestheissueofsmallangularvelocityandprovidesamuchmoredirectpathtoreachthegoal.
Anotherproblemisthelimitedsensingability.Inthispaper,thealgorithmisimplementedonamobilerobotwitha180-degreePOVLRF.