cvpr2013-Modeling Mutual Visibility Relationship in Pedestri(2)

发表在CVPR13上的行人检测方法

halfpartincreaseswhenBob’sright-halfpartisfoundtobevisible.Andthedetectioncon denceofAlicecorrespond-inglyincreases.Inthisexample,thecompatiblerelationshiphelpstodetectAliceinFig.1(b).

2.Incompatiblerelationship.Itmeansthattheocclusionofonepartindicatesthevisibilityoftheotherpart,andviceversa.FortheexampleinFig.1(b),AliceandBobhavesostrongoverlapthatoneoccludestheother.Inthiscase,Alice’sright-halfpartandBob’sleft-halfpartareincom-patiblebecausetheyshallnotbevisiblesimultaneously.IfapedestriandetectordetectsbothAliceandBobwithhighfalsepositiverateinFig.1(b),thenthevisibilitycon denceofAlice’sright-halfpartincreaseswhenBob’sleft-halfpartisfoundtobeinvisible.AndAlice’sdetectioncon denceiscorrespondinglyincreased.Therefore,incompatiblerela-tionshiphelpstodetectAliceinthisexample.

Theseobservationsmotivateustojointlyestimatetheocclusionstatusofco-existingpedestriansbymodelingthemutualvisibilityrelationshipamongtheirparts.Inthispa-per,weproposetolearnthecompatibleandincompatiblerelationshipbyadiscriminativedeepmodel.

Themaincontributionofthispaperistojointlyestimatethevisibilitystatusesofmultiplepedestriansandrecognizeco-existingpedestriansviaamutualvisibilitydeepmodel.Overlappingpartsofco-existingpedestriansareplacedatmultiplelayersinthisdeepmodel.Withthisdeepmodel,1)overlappingpartsatdifferentlayersverifythevisibilityofeachotherformultipletimes;2)thecomplexprobabilisticconnectionsacrosslayersaremodeledwithgoodef ciencyonbothlearningandinference.Thedeepmodelissuitableformodelingthemutualvisibilityrelationshipbecause:1)thehierarchicalstructureofthedeepmodelmatcheswiththemultilayersofthepartsmodel;2)overlappingpartsatdifferentlayersverifythevisibilityofeachotherformul-tipletimesinthedeepmodel;3)thecomplexprobabilisticconnectionsacrosslayersofpartsaremodeledwithgoodef ciencyonbothlearningandinference.Themutualvis-ibilitydeepmodeleffectivelyimprovespedestriandetec-tionperformancewithlessthan5%extracomoputationinthedetectionprocess.ItachievesthelowestaveragemissrateontheCaltech-TraindatasetandtheETHdataset.OnthemorechallengingPETSdatasetlabeledbyus,includ-ingmutualvisibilityleadsto8%improvementonthelow-estaveragemissrate.Furthermore,ourmodeltakespartdetectionscoresasinputanditiscomplementarytomanyexistingpedestrianapproaches.Ithasgood exibilitytointegratewithothertechniques,suchasmorediscrimina-tivefeatures[31],scenegeometricconstraints[27],richerpartmodels[40,38]andcontextualmulti-pedestriandetec-tioninformation[30,26,36]tofurtherimprovetheperfor-mance.

2.RelatedWork

Sincevisibilityestimationisthekeytohandleocclu-sions,manyapproacheswereproposedforestimatingvisi-bilityofparts[2,10,11,32,35,33,29,22,34,21].Wangetal.[32]usedtheblock-wiseHOG+SVMscorestoes-timatevisibilitystatusandcombinedthefull-bodyclas-si erandpart-basedclassi ersbyheuristics.Enzweileretal.[11]estimatedthevisibilityofdifferentpartsus-ingmotion,depthandsegmentationandthencomputedtheclassi cationscorebysummingupmultiplevisibil-ityweightedcuesofparts.Substructureswereusedin[2,10].Eachsubstructurewascomposedofasetofpartdetectors.Andthedetectioncon dencescoreofanob-jectwasdeterminedbytheexistenceofthesesubstructures.TheAnd-Orgraphwasusedin[35]toaccumulatehard-thresholdedpartdetectionscores.Recently,theapproachesin[10,25]utilizedthevisibilityrelationshipamongpartsforisolatedpedestrian.However,thepartvisibilityrela-tionshipamongco-existingpedestrianswasnotexploredin[2,10,11,32,35].Inordertohandleinter-humanocclu-sions,thejointpart-combinationofmultiplehumanswasadoptedin[33,29,22,34,21].Theseapproachesobtainthevisibilitystatusbyocclusionreasoningusing2-Dvis-ibilityscoresin[33,29,22]orusingsegmentationresultsin[34,21].Theymanuallyde nedtheincompatiblere-lationshipamongpartsofmultiplepedestriansthroughtheexclusiveoccupancyofsegmentationregionorpartdetec-tionresponse,whileourapproachlearnstheincompatiblerelationshipfromtrainingdata.Inaddition,thecompatiblerelationshipwasnotusedbytheseapproaches.

Thearticulationrelationshipamongthepartsofmulti-pleobjects,parameterizedbyposition,scale,size,rotation,wasinvestigatedascontext[39,36,37,5].Nearbydetec-tionscoreswasconsideredascontextin[6].Butitdidnotconsiderthevisibilityrelationshipofco-existingpedestri-ans,whichisthefocusofourapproach.Thepartvisibilityrelationshipamongco-existingpedestrianshasnotbeenin-vestigatedyetandiscomplementarytothesecontext-basedapproaches.

Deepmodelhasbeenappliedfordimensionalityreduc-tion[17],handwrittendigitrecognition[16,20,24],ob-jectrecognition[18,20],faceparsing[23],facialexpres-sionrecognitionandscenerecognition[28].Hintonetal.[16]provedthataddinganewlayer,ifdonecorrectly,cre-atesamodelthathasabettervariationallowerboundonthelogprobabilityofthetrainingdatathanthepreviousshal-lowermodel.Bengio[1]provedthatanarchitecturewithinsuf cientdepthcanrequiremanymorecomputationalel-ements,potentiallyexponentiallymore(withrespecttoin-putsize),thanarchitectureswhosedepthismatchedtothetask.Krizhevskyetal.[19]proposedadeepmodelthatachievedstate-of-the-artperformanceforobjectdetectionandrecognitionontheImageNetdataset[4].Recently,deep

百度搜索“77cn”或“免费范文网”即可找到本站免费阅读全部范文。收藏本站方便下次阅读，免费范文网，提供经典小说教育文库cvpr2013-Modeling Mutual Visibility Relationship in Pedestri(2)在线全文阅读。