Abnormal Crowd Behavior Detection Based on the Energy Model(3)

Thereisalsosomeresearchonthreshold-basedmethodsinthevideoanalysis.Oncethetargetvalueexceedbyapresetthreshold,themonitoroutputsanalert.Withmultiplelocalmonitorscollectinglow-levelstatistics(velocityordirection)andusingBayesiansurprise,Xieetal.[10]presentatwolayerthreshold-basedapproachtodetectabnormalevents.Chenetal.[11]proposeatwo-stagehierarchicalclusteringapproachwhichcangroupoptical owsintocrowdsanddetectabnormaleventsusingtheforce eldmodel.Em-ployinganunsupervisedclusteringtechniquetosegmentthevideointospatial-temporalvolumes,Luetal.[12]utilizesspatio-temporalshapeand owcorrelationtodetectactivity.Withoutlearningprocessandtrainingdata,Ihaddadeneetal.[13]ingsocialforcemodel,Mehranetal.[14]introduceamethodtodetectabnormalbehaviorsincrowdscenes.Consideringdensityanddirectionsimultaneously,authorsin[15]usethemotionheatmaptode nearegionofinterestandestimateentropyoftheframestodetectabnormalbehaviors.Adoptingthemotionfeature,Zhongetal.[16]de nethecrowdenergytorepresentcrowddensityanddetectabnormalactivities.Caoetal.[17]

496

combinecrowdkineticenergy,motionvariationanddirectionvariationfortheanomalydetection.

Withouttrainingdata,thethreshold-basedmethodsareeasytobeimplemented.However,thesemethodsalwayshavetodealwithocclusions,trackingandsegmenting.Thethresholdisdif culttobedetermined,alwaysbyexperience,andthefalsealarmrateisusuallyhigh.

III.SYSTEMOVERVIEW

Ourapproachcontributestothesecondcategoryofrelatedworks.However,weavoidtrackingofobjectstoaverttypicalproblemssuchasextensiveclutteranddynamicocclusions.Toourknowledge,thereislittleresearchonanomalydetec-tionconsideringdensity,distributionandmotioninformationsimultaneously.Inthispaper,weadoptanovelmodeltoobtainaccurateestimationofcrowddensitywhichishelpfulandcrucialtodetectionofabnormalactivities.Thesystemin-cludesthreeprimaryparts:Crowddensityestimation,CrowdDistributionIndexandKineticenergy,justasshowninFig.2.Inthe rstpart,asmostexistingvideoanalysis

approaches,

Fig.2.Systemarchitecture

weusetheadaptiveGaussianMixtureModeling(GMM)methodtoextracttheforeground,andthenexcludethenoisebyapplyingthebinarymorphologytechnique.Basedontheimagepotentialenergymodelin[18],wecanobtainaccurateestimationofcrowddensity.Inthesecondpart,webuildtheforegroundhistogramson and axisrespectively,andthencomputetheprobabilitydistribution.DistributionEntropyisde biningtheCrowdDensityandtheCrowdEntropy,wede netheCrowdDistributionIndextodetectpedestraingathering.Inthethirdpart,basedonCrowdDistributionIndex,weimprovethede nitionofkineticenergyin[16],andthemodi edkineticenergymodelcandetectpeoplerunningeffectively.