Coupling Detection and Data Association for Multiple Object Tracking

《Coupling Detection and Data Association for Multiple Object Tracking》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、CouplingDetectionandDataAssociationforMultipleObjectTrackingZhengWu,AshwinThangali,StanSclaroffandMargritBetke∗DepartmentsofComputerScience,BostonUniversity,Boston,MA02215{wuzheng,tvashwin,sclaroff,betke}@cs.bu.eduAbstractalarmsandmisseddetectionsotherwisepropagatetothedataassociatio

2、nmoduleandfalsematchesneedtobecor-Wepresentanovelframeworkformultipleobjecttrack-rectedlater.Incontrast,weshowthaterrorpropagationinginwhichtheproblemsofobjectdetectionanddataas-fromdetectiontodataassociationcanbeavoidedifbothsociationareexpressedbyasingleobjectivefunction.Thetasks,d

3、etectionanddataassociation,arecombinedintoaframeworkfollowstheLagrangedualdecompositionstrat-singlemoduleandsolvedsimultaneouslybyoptimizingaegy,takingadvantageoftheoftencomplementarynaturesingleobjectivefunction.Thiscouplingideaappearsattrac-ofthetwosubproblems.Ourcouplingformulatio

4、navoidstivebutintroducesnewchallengesaswell:1)Whattypeoftheproblemoferrorpropagationfromwhichtraditionalobjectivefunctionshouldbeused?Manyexistingdetectiondetection-trackingapproachestomultipleobjecttrack-methodshavenotevenbeenformalizedwithanobjectiveingsuffer.Wealsoeschewcommonheur

5、isticssuchasnon-function.2)Howcanthenewobjectivefunctionbesolved?maximumsuppressionofhypothesesbymodelingthejointManycurrentdataassociationmethodsarecomplicatedandimagelikelihoodasopposedtoapplyingindependentlike-approximatesolutionstointractableproblems.Anewob-lihoodassumptions.Ourc

6、ouplingalgorithmisguaranteedjectivefunctionthatcouplesdetectionanddataassociationtoconvergeandcanhandlepartialorevencompleteoc-mightbeevenmoredifficulttooptimize.3)Howcanscala-clusions.Furthermore,ourmethoddoesnothaveanyse-bilityoftheproposedmethodbeensured?Computervisionverescalabili

7、tyissuesbutcanprocesshundredsofframessystemsfacedemandsforbeingabletotracklargenum-atthesametime.Ourexperimentsinvolvechallenging,bersofobjectsindenseformations.Givensuchlargeinputnotablydistinctdatasetsanddemonstratethatourmethodsizes,anefficientalgorithmtooptimizethenewobjectivecana

8、chieveresult