Abnormal Crowd Behavior Detection Based on the Energy Model

Proceeding of the IEEE

International Conference on Information and Automation Shenzhen, China June 2011

AbnormalCrowdBehaviorDetection

BasedontheEnergyModel

GuogangXiong ,XinyuWu ,Yen-LunChen ,andYongshengOu

InstitutesofAdvancedTechnologyChineseAcademyofSciencesShenzhen,GuangdongProvince,China

TheChineseUniversityofHongkong,HongKong,China

{gg.xiong,xy.wu,yl.chen,ys.ou}@

Abstract—Inthispaper,wepresentanovelmethodtodetecttwotypicalabnormalactivities:pedestraingatheringandrunning.Themethodisbasedonthepotentialenergyandkineticenergy.Reliableestimationofcrowddensityandcrowddistributionare rstlyintroducedintothedetectionofanomalies.Estimationofcrowddensityisobtainedfromtheimagepotentialenergymodel.Bybuildingtheforegroundhistogramonthe and axisrespectively,theprobabilitydistributionofthehistogramcanbeobtained,andthenwede netheCrowdDistributionIndex( )torepresentthedispersion.TheCrowdDistributionIndex( )isusedtodetectpedestrainsgathering.Thekineticenergyisdeterminedbycomputationofoptical owandCrowdDistributionIndex,andthenusedtodetectpeoplerunning.Thedetectionforabnormalactivitiesisbasedonthethresholdanalysis.Withouttrainingdata,themodelcanrobustlydetectabnormalbehaviorsinlowandmediumcrowddensitywithlowcomputationload.IndexTerms—Intelligentsurveillance,Imagepotentialenergymodel,Abnormalevents,Crowdanalysis.

Shenzhen

objects,suchasbelongingdropping,loiteringandcrossingoverthefence.Asonlyafewpeoplemovinginthescenes,theseapproachescanimplementdetectingandsegmentingeasily.However,whentheenvironmentbecomescompli-cated,asshowninFig.1,thesemethodswillbesubjectedtosevereocclusionswhichmakesthetracking,detectingandsegmentingdif culttoimplement.Basedontheabovefactors,therearefewattemptstomodellargergroupsofpeoplewhichshouldbepaidmoreattention

to.

I.INTRODUCTION

Thedecreasingcostsofvideosurveillanceequipmentshaveresultedinlargevolumesofvideodata.However,thisexcessiveamountofinformationhasnotbeenmetwithenoughhumanoperators[1].Ontheotherhand,techniquesonimageandvideoanalysisdeveloprapidly.Duetotheabovetwofactors,crowdanalysisincomputervisionhasbecomeapopularresearchtopicinnumerouscountries.Modelsabletodetectabnormaleventswithinvideostreamscanservearangeofapplications,suchassecurityautomationsysteminpublic,coalminesurveillanceandintelligentanalysisapplication.Inanysuchcase,automaticalanomalydetectionwouldsigni cantlyimprovetheef ciencyofvideoanalysis,savingvaluablehumanattentionforonlythemostsalientcontent[2].

Mosttraditionalapproachesonanomalydetectionalwaysaimatspeci canomaliesofsinglepersonorafewmoving

workdescribedinthispaperispartiallysupportedbytheNature

ScienceFoundationofChina(61005012),byShenzhen/HongkongInnova-tionCircleProject(ZYB200907070024A)andbythegrantfromShenzhenpublicscienceandtechnology.TheauthorswouldliketothankMr.RuiqingFu,Mr.LeiZhang,Mr.KeXu,andMr.LongHanfortheirvaluablecontributiontothisproject.

This

(a)People

gathering

Fig.1.

(b)Peoplerunning

Typicalabnormalscenes.

Thispaperaimstopresentaneffectivemodeltodetecttwokindsofanomalieswhicharethemostprimaryandprevalentinpublicscenes.Generallyspeaking,pedestriangatheringandrunningisanemergencysignalindicatingsomeabnormaleventshappening,surveillancesystemsshoulddetectthemautomaticallyintime.Therestofthispaperisorganizedasfollows.AsummaryoftherelatedworkisgiveninSection2.OursystemdiagramisdescribedinSection3.Wepresenttheimagepotentialenergymodeltoestimatethecrowddensityinsection4.InSection5,wede netheCrowdDistributionIndex.Modi edde nitionofkineticenergyisgiveninSection6.InSection7,wepresenttheexperimentalresultsondifferentvideoclips.Inthelastsection,wesummarizetheapproachandpresentsomecluesforfutureresearchwork.

II.RELATEDWORK

Abnormalcrowdbehaviordetectioncanbedividedintotwobroadfamiliesofapproachesnamedmachine-learning-basedmethodsandthreshold-basedmethods.

978-1-61284-4577-0270-9/11/$26.00 ©2011 IEEE

495