Abnormal Crowd Behavior Detection Based on the Energy Model(7)

VII.EXPERIMENTALRESULTS

Theabnormaleventsaredetectedinastaticmethodbycomparing and estimationis withaspeci cthreshold.Thethresholdnecessaryforeachcrowdscenebecausethethresholdvariesdependingonthecameraposition.Takepedestraingatheringforanexample,weshowtheprocedureforeachframeindetailasfollows:

IF( >threshold)THENDOTcnt++ELSETcnt=0ENDIF

IFTcnt>10THENDOalarm!ENDIF

Whenasuddenchangeof or noiseandthesystemwould arise,weconsideritcouldbenotgiveanalarmsignal.Whenthe or isgreaterthanitsthresholdfor10consecutiveframes,thesystemalarm.Thisprocedurewillreducethefalsealarmrateeffectively.

ThesystemhasbeenimplementedinC++andtestedusingAMD3500+PCwith994MhzCPUand1GBmemory.Itcanruninnearrealtime(20fps).Inordertoevaluatethepowerofouralgorithm,wechoosefourvideoclipstoshowtheresults.Eachclipconsistsoftwokindsofsituation:normalandabnormal.