Abnormal Crowd Behavior Detection Based on the Energy Model(2)

Machine-learning-basedapproachesareadoptedincrowdanalysismoreandmoreinrecentyear,suchasPrincipalComponentsAnalysis(PCA),K-meansandHiddenMarkovModel(HMM).Kratzetal.[1]constructmotion-patterndistributionswhichcapturethevariationsoflocalspatio-temporalmotionpatternstorepresentthevideovolume,thenderiveadistribution-basedHMM,andimprovetheframeworkbyconstructingacoupledHMM.Authorsin[2][3]adoptedaspatio-temporalMRFmodeltodetectab-normalactivities.Anunsupervisedtechniquefordetectingabnormalbehaviorinlargevideosetsispresentedin[4],rathertobuildexplicitmodelofnormalevents,ingoptical owpatternstoestimatethemodelparameters,Andradeetal.[5]adoptHMMstorepresentthe owsequences.Basedonadynamicconditionrandom eld(DCRF)model,Yinetal.[6]proposeanewSpatio-TemporalEventDetectionalgorithm.ByemployingPrincipalComponentAnalysisforfeatureselection,Wuet.al.[7]adoptSupportVectorMachinetoclassifyhumanbehaviors.Bydividingwholeframeintosmallblocks,Wangetal.[8]useKLTcornerstorepresentmovingobjectsandclustermotionpatternsinanunsupervisedway.Wuetal.[9]extractthechaoticdynamicsofallrepresentativetrajectories,andthenusetheprobabilisticmodeltotrainthesechaoticfeaturesets.Themachine-learning-basedmethodcanbequiteeffec-tiveintheenvironmentwhere“normal”activityiswell-de nedandconstrained.However,asthenumberofdifferent“normal”observedactivitytypescaneasilysurpassthatofunusualtypes,de ningandmodelingthe“normal”activityinanunconstrainedenvironmentisalwaysimpossible.