文稿background subtraction via cvpr paperxin背景减法.pdfVIP

BackgroundSubtractionviaGeneralizedFusedLassoForegroundModeling

BoXinYuanTianYizhouWangWenGao

Nat’lEngineeringLaboratoryforTechnology

CooperativeMedianetInnovationCenter

KeyLaboratoryofMachinePerception(MoE)

Sch’lofEECS,PekingUniversity,Beijing,100871,

[21,27,12]andneuralnetworkmodels[9,17]etc.Non-

parametricmodelsarealsoproposedforimprovedeffi-

BackgroundSubtraction(BS)isoneofthekeystepsciency[1].Thepixel-wisemodelsarepronetoresulting

inanalysis.Manybackgroundmodelshavebeeninfragmentaryforegrounds,i.e.thereareboth“holes”in

proposedandachievedpromisingperformanceonpublictheforegroundsandfalsepositivepixelsfromtheback-

datasets.However,duetochallengessuchasilluminationground.Whereas,themodelsofthesecondcategory,i.e.

change,dynamicbackgroundetc.theresultedforegroundframe-wisemodels,usuallyachievterperformanceby

segmentationoftenconsistsofholesaswellasbackgroundexploringstructureinformationofthebackground.These

noise.Inthisregard,weconsidergeneralizedfusedlassoworkscanbegenerallyviewedasfollow-upsofthecele-

regularizationtoquestforintactstructuredforegrounds.bratedeigen-backgroundmodelproposedin[20],ofwhich

Togetherwithcertainassumptionsaboutthebackground,thekeyassumptionisthatwhencameramotionissm