Experiential sampling based foregroundbackground segmentation for video surveillance

Segmentation of foreground and background has been an important research problem arising out of many applications including video surveillance. A method commonly used for segmentation is “background subtraction ” or thresholding the difference between th

EXPERIENTIALSAMPLINGBASEDFOREGROUND/BACKGROUNDSEGMENTATION

FORVIDEOSURVEILLANCE

PradeepK.Atrey ,VinayKumar ,AnuragKumar andMohanS.Kankanhalli

SchoolofComputing,NationalUniversityofSingapore

IndianInstituteofTechnology,Kharagpur,India

theprocessingisdoneonlyontheattentionsamplesinsteadoftheentiredata.TheEStechniquehasbeenshownuseful

Segmentationofforegroundandbackgroundhasbeenanim-inmanyapplicationsincludingfacedetectionandmonologue

portantresearchproblemarisingoutofmanyapplicationsin-detectioninvideo[1].Wealsoexploitthetemporalredun-cludingvideosurveillance.Amethodcommonlyusedfor

dancyofthevideotoreducethenumberofcomputations.

segmentationis“backgroundsubtraction”orthresholdingthe

WehaveusedadaptiveGaussianmethodtomodelthe

differencebetweentheestimatedbackgroundimageandcur-backgroundasdescribedbyStaufferetal.[2]andfurther

rentimage.AdaptiveGaussianmixturebasedbackground

improvedbyKaewTraKulPongetal.[3].Thesemethodsdo

modellinghasbeenproposedbymanyresearchersforincreas-thecomputationsonthewholeimagewithouttakingintocon-ingtherobustnessagainstenvironmentalchanges.However,

siderationoftheregionsofinterest.Itconsumesasigni cant

allthesemethods,beingcomputationallyintensive,needtobe

amountoftimeindoingunnecessarycomputationsespecially

optimizedforef cientandreal-timeperformanceespecially

innon-busyenvironmentswheremostoftheframescaptured

atahigherimageresolution.Inthispaper,weproposeanim-bythecamerahasaclearbackgroundandshouldnotbegiven

provedforeground/backgroundsegmentationmethodwhich

muchattention.TheproposedmethodthatintegratesEStech-usesExperientialSamplingtechniquetorestrictthecomputa-niquewiththealreadyproposedmethodsofbackgroundseg-tionaleffortsintheregionofinterest.Weexploitthefactthat

mentationshowsasigni cantimprovementinthecomputa-theregionofinterestingeneralispresentonlyinasmallpart

tionalspeed.Thisimprovementinspeedisachievedatthe

oftheimage,therefore,theattentionshouldonlybefocused

costofminorlossinaccuracy.Thislosshoweverisaccept-inthoseregions.Theproposedmethodshowsasigni cant

ableinlightofthefactthatanyeventlastsforsuf ciently

gaininprocessingspeedattheexpenseofminorlossinac-largenumberofvideoframesandthenumberofvideoframes

curacy.Weprovideexperimentalresultsanddetailedanalysis

inwhichtheforegroundismissed(inourmethod)isveryless.

toshowtheutilityofourmethod.

And,ofcourse,nosurveillancetaskendsupatthesegmen-tationofforegroundonly,ratheritundergoesfurtheranalysisvizeventsdetection,monitoringandtrackingetcwhichrelies1.INTRODUCTION

onaseriesofvideoframesbeforeconcludingaboutanevent.HenceeveniftheforegroundinafewframesaremissedbyReal-timeprocessingofvideohasalwaysbeenaproblemin

thedetector,itdoesnotaffectthe nalobjectiveappreciably.manyapplicationsincludingautomaticvideosurveillance.In

Wehaveshownthroughexperimentsthatthelossinaccuracyautomaticvideosurveillance,oneofthemajorstepsinvideo

basedhuman-activityrecognitionistheforeground/backgroundisverysmallcomparedtothegainincomputationalspeed.

Ourcontributionsinthispaperaresummarizedasfollows.segmentationwhichtakessubstantialamountofcomputation

WehaveproposedanExperientialSamplingtechniquebasedtime.Inthispaper,wefocusonimprovingthecomputational

foreground/backgroundsegmentationmethodwhichprovidesef ciencyofanexistingforeground/backgroundsegmenta-improvedcomputationalef ciencyatthecostofnegligibletionalgorithmtomeetthereal-timerequirements.

lossinaccuracy.ThecoreideaofourmethodistouseExperientialSam-pling(ES)technique[1]to ndtheregionofattentionineach

videoframeandtorestricttheprocessingtoit.TheEStech-2.RELATEDWORK

niqueutilizesthepastexperiencetomodelthegoalbasedcontextualattentionusingwhichit ndstheregionwheretheSinceweusebothforeground/backgroundsegmentationascomputationsneedtobedone.Thegoal,inourcase,istoseg-wellasexperientialsampling,wedescribetherelatedworks

menttheforegroundfrombackground.TheEStechniquepro-inbothofthem.Backgroundsubtractioninvolvesmodelling

videsanef cientwaytoderivetheattentionsamplesfromtheareferenceframe,subtractingthecurrentframe,andthenmedia(sensor)samples.Oncewehavetheattentionsamples,thresholdingtheresult.Thismodellingthoughissimplebut

ABSTRACT

Experiential sampling based foregroundbackground segmentation for video surveillance相关文档

最新文档

返回顶部