The rsttwovideoclipsareobtainedbyastationarycamerainstalledonthebuildingBofShenzhenInstituteofAdvancedTechnologyinatopdownview.Fig.7andFig.8showtheresultandanalysis.Theoriginalvideoframesizeis320*240.Underthiscircumstance,thethresholdof is6andthethresholdof is
180.
(a)Typicalscreen-capturesofthevideo
clip
(b) curveFig.7.
Pedestrainsgathering.
Inthe rstclipshowninFig.7,peoplearegatheringfromdifferentdirectionsandthenwalktogetherinonedirection,
499
ouralgorithmcorrectlydetectstheanomalyintime.Fig.7(b)showsthecurveof ,theabnormalgatheringoccursatthe218thframe.Asthereisnopeoplerunninginthescene,curveof is
ignored.
(a)Typicalscreen-capturesofthevideo
clip
(b) curve
Fig.8.
Pedestrainsrunning.
InthesecondclipshowninFig.8,peoplewalkinalowspeed,andthenbegintorunsuddenly.Theabnormalrunningisdetectedatthe500thframe.Fig.8(b)showsthecurveof .Nopeoplegatheringinthisclip,sothecurveof isabandoned.InFig.8(b),wecanseeasharppeakwhichisactuallynoiseoccursataboutthe400thframe;however,itonlylast6frames,oursystemavoidsfalsealarmsuccessfully.Anothertwovideoclipsdownloadedfromtheinternetarewithhighcrowddensity,andtwotypicalanomaliesoccurinthesameclip.Theframesizeis768*576.Underthiscircumstance,thethresholdof is12andthethresholdof is40000.Fig.9andFig.10showtheexperimentalresults.
InthethirdclipshowninFig.9,wecanseethecrowddensityishighandocclusionissevere.However,ouralgo-rithmdetectthepedestraingatheringinalocalregionat100thframeeffectively.InFig.9(a), isred,whichindicatesthatpeoplegatheringistakingplace.Fig.9(b)presentthecurveof .Inthesameclip,aftergathering,peoplestarttorunsuddenly.Fig.9(c)showsthecurveof detectedat355thframe.
,whererunningisAsshowninFig.10,atframe44th,ouralgorithmdetectabnormalactivitiestakingplace,while and red,whichmeanspedestrainsgatheringandrunning arebothoccursimultaneously.
VIII.CONCLUSIONS
AND
FUTUREWORK
Inthispaper,wepresentanovelmethodtodetecttwotypicalabnormalactivities:pedestraingatheringandrunning.Withouttrainingdata,themodelcanrobustlydetectabnormal

