zhangyli8nevatia uscedu Abstract We propose a network 64258ow based optimization method for data association needed for multiple object tracking The maximumaposteriori MAP data association prob lem is mapped into a cost64258ow network with a nonoverl ID: 33403
Download Pdf The PPT/PDF document "Global Data Association for MultiObject ..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
GlobalDataAssociationforMulti-ObjectTrackingUsingNetworkFlowsLiZhang,YuanLiandRamakantNevatiaUniversityofSouthernCaliforniaInstituteofRoboticsandIntelligentSystemsWeproposeanetworkowbasedoptimizationmethodfordataassociationneededformultipleobjecttracking.Themaximum-a-posteriori(MAP)dataassociationprob-lemismappedintoacost-ownetworkwithanon-overlapconstraintontrajectories.Theoptimaldataassociationisfoundbyamin-costowalgorithminthenetwork.The Figure1.Detectioninputandtrackingresult:ourmethodcanre-movefalsealarms,recovertrajectoriesandinfereventssuchasmisseddetectionsandocclusions.Inourapproach,dataassociationisdenedasaMAPes-timationproblemgivenasetofobjectdetectionresultsas 2.RelatedworkTotrackmultipleobjects,oneapproachistomakedataassociationdecisionsframe-by-frame(orinasmalltimewindow)asin[].Whilesuchmethodshaveshownverygoodperformance,consideringmoreframesbeforemakingassociationdecisionsshouldgenerallyhelpbetterovercomeambiguitiescausedbylonger-termocclusionsandfalseormisseddetections.Manyglobalapproachesthatusemoreinformationhavebeenexploredtoovercomeerrorsofdetections.Onestrat-egyistooptimizeonetrajectoryatatimethroughtheen-tiresequence;thishasbeenusedinDynamicProgrammingbasedmethods,suchas[].Greedystrategiesarethenusedtocombinethetrajectoriesandhandlepotentialcon-icts.Itisdifcultforthesemethodstomodelocclusionsbecausetrajectoriesareoptimizedseparately.Anotherap-proachistooptimizemultipletrajectoriessimultaneously;multi-HypothesisTracking(MHT)[]andJointProbabilis-ticDataAssociationFilters(JPDAF)[]aretworepresenta-tiveexamples.Alsoin[],detectionandestimationoftra-jectoryhypothesesarecoupledbyQuadraticBooleanPro-gramming.Asthehypothesessearchspaceiscombinato-rial,suchmethodscanonlyoptimizeoveralimitedtimewindow,andhypothesesmuststillbepruned.SamplingmethodssuchasMCMC[]havealsobeenemployedtondapproximatesolutions.Occlusionsareusuallymodeledasmergingandsplittingoftrajectoriesinthesemethods.TrackletStitching[]andLinearProgramming(LP)basedtracking[]aretwootherapproachesseekingtoop-timizealltrajectoriessimultaneouslyovertheentirese-quence.[]rstgeneratestracklets,whicharefragmentsoftracksformedbyconservativegroupingofdetectionre-sponses.ThetrackletsarethenconnectedbyHungarianpartitioningalgorithm.Thismethodassumesalltrackletstocorrespondtotrueobjecttrajectoriesandhenceishardtoextendtorawdetectionsineachframewheremanyfalsealarmsarelikelytobepresent.[]buildsasetofsubgraphsforeveryobjecttrajectorywithedgesbetweenthemrepre-sentingtheobjectinteractions.Amulti-pathsearchproblemonthesubgraphsisthensolvedapproximatelybylinearpro-grammingandrounding.Itassumesinter-objectpositionstoberelativelystable,andthenumberoftargettobexed.3.OurapproachWedenedataassociationasaMAPproblem.Theprob-lemisthenmappedintoacost-ownetwork,andsolvedwithamin-costowalgorithm.Themappingisbasedontheobservationthatthereisananalogybetweennd-ingnon-overlappingobjecttrajectoriesandndingedge-disjointpathsinagraph;thelattercanbesolvedefcientlybynetworkowalgorithms.Werstpresenttheformula-tion,andthenprovidethemin-costowsolution.3.1.MAPundernon-overlapconstraintsbeasetofobjectobservations,eachofwhichisadetectionresponse,,whereistheposition,isthescale,istheappearanceandisthetimestep(frameindex)oftheobject.Asingletrajectoryhypothesisisdenedasanorderedlistofobjectobservations,.Anassociationhypothesisisdenedasasetofsingletrajectoryhypotheses,Theobjectiveofdataassociationistomaximizethepos-terioriprobabilityofgiventheobservationsetT|XX|TassumingthatthelikelihoodprobabilitiesareconditionallyindependentgiventhehypothesisItisdifculttooptimizeEqn.directly,becausethespaceofishuge.However,wecanreducethesizeofthesearchspacebyusingtheobservationthatoneobjectcanonlybelongtoonetrajectory.Thistranslatesintotheconstraintthatcannotoverlapwitheachother,Ifwefurtherassumethatmotionofeachobjectisindepen-dent,wecandecomposeEqn.ThetermsinEqn.()aredenedasfollows:isthelikelihoodfunctionofobservationBernoullidistributionisusedtomodelthecasesofanobser-vationbeingatruedetectionaswellasbeingafalsealarmistheprobabilityforbeingafalsealarm).modeledasaMarkovchain,whichincludesinitialization,terminationprobability,andtran-sitionprobabilities.ThepreciseformofthesefunctionsandtheirestimationfromtrainingdataaredescribedlaterinSection Notethatthelikelihoodfunctioncanmodelnotonlytheobservationsthatareassociatedin.truede-tections,butalsothosethatarenotassociated,.falsealarms.Thisallowsthemethodtoselectobservations,ratherthanassumealltheinputstobetruedetections,with-outadditionalprocessingtoremovefalsetrajectoriesafter3.2.Min-costowsolutionTocouplethenon-overlapconstraintswiththeobjectivefunction,thefollowing0-1indicatorvariablesaredenedstartsfromendsatisrightafterItseasytoseethatthesevariablesaredeterminedforagivenassociationhypothesis,andviceversa.isnon-overlapifandonlyifNext,weincorporateindicatorsinlogarithmoftheob-jectivefunction,log(1subjecttoEqn.,where=log Thisformulationcanbemappedintoacost-ownetworkwithsourceandsink.Givenanobservationset "# $ $! & ' () $Figure2.Aexampleofthecost-ownetworkwith3timestepsand9observations:foreveryobservation,createtwonodescreateanarcwithcostandow,anarcs,uwithcosts,uandows,u,andanarcwithcostandow.Foreverytran-,createanarcwithcostandow.AnexampleofsuchagraphisshowninFigure.Eqn.isequivalenttotheowconservationconstraintandEqn.tothecostofowin.Findingoptimalassociationhypothesisequivalenttosendingtheowfromsourcetosinkminimizesthecost.Thecost-ownetworkformulationisanintuitiverepre-sentationofmultipleobjecttracking:eachowpathcanbeinterpretedasanobjecttrajectory,theamountoftheowsentfromisequaltothenumberofobjecttrajectories,andthetotalcostoftheowoncorrespondstothelog-likelihoodoftheassociationhypothesis.Theowconser-vationconstraintguaranteesthatnoowpathsshareacom-monedge,andthereforenotrajectoriesoverlap.Ifalltheedgecostsinwerepositive,themin-costowwouldbethetrivialemptyzero-costow.However,foranyobserva-thatismorelikelytobeatruedetection(thecostofedgeisnegative;thisallowstheop-timalcosttobecomebelowzerobysendingowsthroughthesenegative-costedges.Theoptimalcostshouldbecalculatedoverallpossible,whereistheamountofowsentfromsourcetosink.Itisknownthatforagiven,theminimalcostcanbesolvedforinpolynomialtimebyamin-costowalgorithm[].Theentireoptimizationprocessisde-scribedasAlgorithm1.Itcanalsobeproventhatthemin-imalcostisaconvexfunctionw.r.t.Hencetheenu-merationoverallpossiblecanbereplacedbyaFi-bonaccisearch,whichndstheglobalminimalcostbyatexecutionsofthemin-costowalgorithm.|X|bethenumberofedgesin,which ConstructthegraphV,E,C,ffromobservationsetStartwithemptyowWHILE(canbeaugmented)byone.Findthemincostowbythealgorithmof[IF(currentmincostglobaloptimalcost)Storecurrentmin-costassignmentasglobaloptimum.Returntheglobaloptimalowasthebestassociationhypothesis Algorithm1:FindMAPtrajectoriesbymin-costow.growslinearlywith.Letbethenumberofexecu-tionsofthemin-costowalgorithm,whichisboundedlog(.Therunningtimeofourmethodisthecomplexityofthemin-costowalgorithm.Oneef-cientmin-costowalgorithmisthescalingpush-relabelmethodproposedbyGoldberg[](weusetationfromAndrewGoldbergsNetworkOptimizationLi-braryathttp://www.avglab.com/adrew/soft.html).Thisal-gorithmhasaworst-caserunningtimeof,butusuallytakesmuchlesstimeonrealdata.Wendthattheruntimegrowsonlylinearlywiththenumberofobserva-tionsasshowninSection;likelybecauseofthenatureofthenetworkwheretransitionsbetweenobservationsaretemporallyconstrained.Algorithm1providesageneralframeworkfordataasso-ciation.Differentfrommethodswhicheitheroptimizeeachtrajectoryseparatelyorsufferfromthecombinatorialexplo-sionofthehypothesesspace,thismethodisabletondtheglobaloptimumefciently.Next,weextendourmethodtotrackthroughlong-termocclusions.4.Explicitocclusionmodel(EOM)TheformulationinSectioniscapableoftrackingshort-termmisseddetections,includingthosecausedbyocclu-sion.However,long-termocclusions,ifjusttreatedasmissingdata,cannotbehandledwithoutimpairingperfor-mance.Ifweallowassociationofobservationswithalargetemporalgapbetweenthem,thepossibilityofcreatinger-rantassociationalsoincreases.Toeffectivelytrackthroughlong-termocclusions,weproposetoreasonexplicitlyaboutwhichobjectsmaybeoccludingwhichothersbyconstruct-inganExplicitOcclusionModel(EOM).TheEOMgener-atesasetofocclusionhypothesesandcombinesthemwiththeinputobservationsbyasetofocclusionconstraints.Onlyocclusionsbetweentrackedtargetsareaddressed. ))))51x , //// ////51 f Figure3.Anexampleofaddingoccludedobjecthypothesiscorrespondingchangesinthecost-ownetwork:thesolidrectan-glesareinputobservations;thedashed-linerectangleisancludedobjecthypothesis4.1.OcclusionhypothesesandconstraintsTherststepofEOMistoexpandtheobservationsetbyaddingoccludedobjecthypothesisWesaythatobservationdirectlyoccludableandonlyifthedistanceandscaledifferencearebelowcertainthresholds,anddenethecorrespondingoccludedobjecthypothesisarethepositionandtimestepof,andarethesizeandappearanceofForeverypair)insuchthat)isdirectlyoccludable,generateanewcludedobjecthypothesisandaddittotheobservation.Repeatthisuntilnonewhypothesescanbegener-ated.Theprocedureofaddinghypothesesisillustratedinbetheobservationsetobtainedbyexpandingtheasabove.NotethatX||X|afteralldu-plicatehypothesesareremoved;thenumberofobservationdoesnotgrowexponentially.Inpractice,hypothesesthataresimilarinappearanceandsizearemergedbymean-shiftclusteringtofurtherreducethesizeofInthesecondstepofEOM,theproposedMAPformu-)isappliedagaintotheset,exceptfortwodifferences:rst,thereisnoobservationlikelihoodtermforanyhypothesis;second,asetofocclusionconstraintsareimposedasistheindicatorfunctionforthehypothesisAsanycanbeonly0or1,theseconstraintsguarantee thatanoccludedobjecthypothesiscanbeused()inassociationonlyiftheobjectthatoccludesitisalsousedSincetheobjectivefunctionremainsunchangedthroughtheextension,wecanstilloptimizeEqn.,subjecttoandthenewconstraintEqn..WesolvetheEOM-baseddataassociationbyusinganiterativeapproachbuiltonAlgorithm1.4.2.AniterativesolutionAlgorithm1providesanoptimalsolutionwhenobjectsarenotoccluded.Toaccountforocclusions,wetakethetra-jectoriesfoundbytheAlgorithm1tobetruetrajectoriesandaddhypothesesthatareoccludedbythesetruetrajectoriestothenetwork.Algorithm1isthenappliedtotheexpandednetwork.Thisprocessisrepeatedtoinferocclusionsandassociationsiteratively.Moreprecisely,rsttheoriginalMAPformulationwiththeinputobservationsetissolvedwiththeAlgorithm1.betheoptimalassociationhypoth-esis.Foranyobservation,theindicatorisxedtobe1.Thenoccludedobjecthypothesisisgen-eratedforanythatisdirectlyoccludable.LettheexpandedobservationsetbeX {.Becauseareboundto1,theocclusionconstraintsforanyholdau-tomatically.Therefore,insteadofhavingasetofocclusionconstraints,wenowhaveisequivalenttoasetoflowerboundconstraintsonthevaluesofinthecost-ownetwork.Sincethemin-costowwithlowerboundconstraintscanstillbesolvedbythesamealgorithm[],Algorithm1canbeap-pliedagaintosolvetheoptimaldataassociationonconstraintsEqn..Theprocedureofestimatingmin-costowandexpandingobservationsetisrepeateduntilconver-genceorapresetmaximumnumberofiterationisreached.TheapproachisdescribedasAlgorithm2inTable.Inpractice,wendthatthealgorithmusuallyachievesitsop-timalperformanceaftertwoiterations.OcclusioneventscanbeinferredfromtheoutputofAl-gorithm2whenanoccludedobjecthypothesisisusedinthesolution,whileamisseddetectionisinferredwhenthereistemporaldiscontinuityintheassociation.Basedonthis,wecaninfereventsofocclusionandmisseddetectionasshowninourresults(FigureDifferentfrompreviousworkssuchas[],whichmodelocclusionthroughsplittingandmergingofthetra-jectories,ourmethodgeneratesocclusionhypothesesex-plicitlytorecovertheobservationsthatismissingduetoocclusions,andthereforegivesamoreuniedapproachbe-causerecoveredobservationsaretreatedinthesamewayasinputobservations. betheinputobservationsetLetlowerboundconstraintsetSolveusingAlgorithm1withconstraintForeachforanydirectlyoccludableXX{WHILEnot(convergedormaxnumberofiterationreached)Returnthenaloptimalowassignment Algorithm2:FindMAPtrajectorieswithocclusionreasoning.5.ImplementationdetailsInthissection,wedescribetheestimationofthefourinourframework.Astheyaredirectlyrelatedtotheinputobservations,theycanbeestimatedfromthetrainingdatastatistically.aredenedasmissdetectionrateofthedetectornumberoftrajectories numberofhypothesesAsthenumberoftrajectoriesisdatadependent,anEMap-proachisusedtoestimateduringtheopti-ThemodelofisdenedtoemploytheinformationfromobservationsbyThetermsontherightcorrespondtosize,position,appear-anceandtimegaprespectively,whereconditionalindepen-denceoftheothertermsisassumedgiventimeinterval,exceptbetweenscaleandposition.Thepositionandsizetermsareassumedtobenormaldistributions;theyarelearnedfromtrainingdata.Fortheappearanceterm,twoRGBhistograms,areextractedfromthedetectionresponsestively,andisdenedbasedontheBhattacharyya arethenormaldistri-butionsofbetweenthesameobjectanddifferentobjectsrespectively;theyarelearnedfromtrainingdata. Thetimegapcomponentisdenedbyanexponentialmodelbasedonthemissingrateofthedetectorasisthemaximalallowedtimegap.6.EvaluationsInthissection,weshowresultsofourmethodontwodatasets:theCAVIARvideos[]andtheETHMobileScene(ETHMS)[].Bothdatasetareverychallengingbe-causeoftheheavyocclusionsandpoorimagecontrastfrombackground.Ourmethodisevaluatedbyitstrackingper-formance,detectionperformanceandspeed.OurmethodiscomparedwithWuetal.s[]onCAVIAR,astheyhavere-portedthebesttrackingresultonthedataset.OnETHMSitiscomparedwithEssetal.sdetectionmethod[],asitistheonlymethodevaluatedonthedataset.Theresultsshowgoodperformance,withfewerfalsealarmsandtrajectoryfragmentsthanthepreviousmethods.Ourmethodisalsomoreefcientcomparedtootherglobalmethods.6.1.ExperimentsettingsTheCAVIARdatasetincludes26videosequencesofawalkwayinashoppingcentertakenbyasinglecamerawithframesizeofandframerateof25fps.TheETHMSdatasetincludes4videosequencesofstreetscenestakenbyamovingcamera,withframesizeofframerateof15fps.Bothdatasetincludemanyinter-personocclusionsincrowdedscenes,withpoorcontrastbetweenobjectsandbackground.ForCAVIAR,weteston20videos(25587framestotal)andusetheother6videosfortraining.ForETHMS,wetestonsequence#1(999frames)andusetheother3videosfortraining.TheinputobservationsetisfromtheoutputlesofthehumandetectorbyWuetalal2].However,wedonotmakeuseofthepart-basedreason-ingproposedin[],buttakeallthedetectionresponsesinthelesasourinputset.Peoplethataretoosmallintheimages(lessthan24pixelinwidth)orpartiallyoutofthescenearenotcountedintheevaluation.6.2.TrackingPerformanceWeevaluatethetrackingperformanceaccordingtothefollowingvemetricsproposedin[mostlytrackedtrajectories(MT),thenumberoftrajec-toriesthataresuccessfullytrackedformorethan80%;mostlylosttrajectories(ML),thenumberoftrajecto-riesthataretrackedforlessthan20%;partiallytrackedtrajectories(PT),thenumberoftrajec-toriesthataretrackedbetween20%and80%;fragmentation(FRMT),thenumberoftimesatrajec-toryisinterrupted;IDswitches(IDS),thenumberoftimestwotrajectoriesswitchtheirIDs.TheresultsonCAVIARareshowninTable,whereGTisthenumberoftrajectoriesinthegroundtruth. GT MT PT ML FRMT IDS etal 140 106 25 9 35 17 Algorithm1 140 104 29 7 58 7 Algorithm2 140 120 15 5 20 15 Table1.ComparisonofthetrackingresultsonCAVIARdatasetComparedtotheresultin[],Algorithm1outputsmorePTandmoreFRMTbecauseithasnoocclusionmodel,whileEOM-basedAlgorithm2connectsalargeportionofthetrajectoryfragmentstoyieldmoreMT.SomepictorialresultsaregiveninFiguretoshowthattheEOM-basedAlgorithm2canrecovertrajectoriesfromfullocclusions.InFigure,Object#4isoccludedbyobject#5foralongtimebetweenframe301and389,butisstilltrackedbyAl-gorithm2.Resultimagesalsoshowourmethodprovidereasoningofocclusionsandmisseddetectionsexplicitly.showsthatourmethodcanrecovertrajectoriesinsomecomplicatedcaseswithmanyfullocclusions.6.3.DetectionPerformanceWecomparethedetectionrateandfalsealarmsofourmethodwiththeinputobservationsetandpreviousresultsresults2,14]inTable.ItshowsthatAlgorithm1increasesthedetectionrateandreducesfalsealarmssignicantlycom-paredtotheinputobservationset;Algorithm2furtherim-provesthedetectionratebutwithaslightlyhigherfalsealarmrate.TheimprovementofAlgorithm2ondetectionrateisnotsignicantbecausethenumberoffullyoccludedhumansintheentiretestsetisrelativelysmall.BothAlgo-rithm1and2givebigimprovementsinthefalsealarmratescomparedtothepreviousmethods,whilethedetectionratesaresimilar.Becauseourmethoddoesnotuseadditionalwaystolldetectiongapssuchasthemean-shifttrackerinin2],itcannotrecoveratrajectoryifanobjectisnotdetectedformanyconsecutiveframes.6.4.SpeedWehavemeasuredexecutionspeedofthemethodonsomeCAVIARvideoswhichtypicallyhaveseveralobjectstobetrackedineachframe,thetrajectoriesareusuallylongaspeoplewalkalongthecorridorandtherearepersistentocclusions.Theprocessingtimeoftheobjectdetectorisnotcountedhere.Eventhoughthetheoreticalcomplexityforthemin-costowalgorithmispolynomial,wendthat 0 0 0 0 Figure4.DetectioninputsandtrackingresultsonCAVIAR[]:ourmethodcanremovefalsealarmssuchasinframe301,recovermisseddetectionsuchasinframe258,andtrackthroughheavyocclusionssuchasinframe301and389. Dataset Method Detectionrate FalseAlarmperframe CAVIAR Inputobservationset 72.8% 0.270 etal 75.2% 0.281 Algorithm1 74.3% 0.081 Algorithm2 76.4% 0.105 ETHMS Inputobservationset 64.3% 1.54 etal 47% 1.5 Algorithm1 68.3% 0.85 Algorithm2 70.4% 0.97 Table2.DetectionandfalseAlarmrateonCAVIARdatasetthecomplexitygrowsonlylinearlywiththenumberofob-servationsinAlgorithm2.Inoneexample,thereare7000inputobservationsoverasequenceof3500frames,.140seconds,Algorithm1ndstheglobaloptimumin30sec-ondsusinga3.7GHzPC.Algorithm2,appliedtothesamedata,expandstheinputsetwith11000occludedobjecthy-potheses,andndsasolutioninabout2minutes.Henceourmethodisreal-timeandthetotaltimeislikelytobedominatedbythedetectionstep.Becauseoftheefciencyofouralgorithms,weprocesseveryvideoinCAVIARdatasetglobally,withoutpartition-ingorusingaslidingwindowaswouldbenecessaryforthepreviouscombinatorialalgorithms.Slidingwindowtech-niquesmaystillbeusefulifmuchlongersequencesaretobeprocessedoronlineresultsarerequired.7.ConclusionWehavepresentedanoveldataassociationframeworkformultipleobjecttrackingthatoptimizestheassociationgloballyusingalltheobservationsfromtheentiresequence.Falsealarms,initializationandterminationofthetrajectory,andinferenceofocclusionsismodeledintrinsicallyinthemethod.Anoptimalsolutionisprovidedbasedonthemin-costnetworkowalgorithms.Thoughthecomplexityofthealgorithmsispolynomial,inpractice,wendthemtobehighlyefcient.Experimentresultsindicatethatglobaldataassociationishelpful,especiallyforreducingtrajec-toryfragmentsandimprovingtrajectoryconsistency,whilemaintainingefciency.Theframeworkisgeneralandcanbeeasilyadaptedtoapplytotrackinganyclassofobjectsforwhichreasonabledetectorsareavailable.8.AcknowledgementThisresearchwassupported,inpart,bytheOfceofNavalResearchunderContract#N00014-06-1-0470.TheviewsexpressedheredonotnecessarilyreectthepositionorthepolicyoftheUnitedStatesGovernment.ReferencesencesZ.Khan,T.Balch,andF.Dellaert,MCMC-basedparti-clelteringfortrackingavariablenumberofinteractingtar-PAMI,vol.27,no.11,2005.2005.B.WuandR.Nevatia,Trackingofmultiple,partiallyoc-cludedhumansbasedonstaticbodypartdetection,detection,D.B.Reid,Analgorithmfortrackingmultipletargets,Trans.onAutomaticControl,1979.1979.Y.Bar-Shalom,T.Fortmann,andM.Scheffe,Jointprob-abilisticdataassociationformultipletargetsinclutter,inInformationSciencesandSystems,1980.1980.J.Berclaz,F.Fleuret,andP.Fua,Robustpeopletrackingwithglobaltrajectoryoptimization,in,2006. 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 Figure5.FirstandsecondrowarethedetectionandtrackingonETHMS[];fthtoseventhrowaretrackingonCAVIAR[VIAR[L.Zhang,B.Wu,andR.Nevatia,Detectionandtrackingofmultiplehumanswithextensiveposearticulation,ininH.Jiang,S.Fels,andJ.J.Little,Alinearprogrammingapproachformultipleobjecttracking,,2007.2007.A.G.AmithaPerera,C.Srinivas,A.Hoogs,G.Brooksby,andW.Hu,Multi-objecttrackingthroughsimultaneouslongocclusionsandsplit-mergeconditions,,2006.2006.Q.Yu,G.Medioni,andI.Cohen,Multipletargettrackingusingspatio-temporalmarkovchainmontecarlodataassoci-,2007.2007.B.Leibe,K.Schindler,andL.VanGool,Coupleddetectionandtrajectoryestimationformulti-objecttracking,tracking,R.Ahuja,T.Magnanti,andJ.Orlin,NetworkFlows:Theory,Algorithms,andApplications,PrenticeHall,1993.1993.A.V.Goldberg,Anefcientimplementationofascalingminimum-costowalgorithm,J.Algorithms,1997.1997.http://homepages.inf.ed.ac.uk/rbf/CAVIARDATA1.6,7,8[14]A.Ess,B.Leibe,andL.VanGool,Depthandappearanceformobilesceneanalysis,in,2007.