基于视觉和几何特征的街景图像语义分割-通信与信息系统专业论文.docxVIP

SemanticSegmentationofStreetScenesbyAppearanceandGeometryFeaturesABSTRACTDigitalImageshavebeenwidelyusedinourlifeascarriersofinformation,andoneofthegrandgoalsofcomputervisionistointerpretascenesemanticallygivenaninputimage.Streetscenesarecloselyrelatedtopeople’sdailylife,buttheseimagesarecomplicatedtoprocess,challengesandopportunitiesexistatthesametimetosemanticallysegmentsuchimageswithgoodperformance.Currentsegmentationalgorithmscouldonlydealwithimagesofsomeapplicationfield,whichcontainafewclasses.Besides,generalimagesegmentationalgorithmsaresensitivetothechangesofrotation,scaleandintensityandsoon,whichoftenfailtoproducesemanticregionsidenticaltotheobjects.Recently,combiningobjectsegmentationandrecognitionhasbeenparticularlyactiveandnotableprogresshasbeenmade,butchallengesremainforgettingfineobjectboundaries,whenlargenumberofclassesoccursimultaneouslyandobjectsvarydramaticallyinsizeandshape.Onthecontrary,streetscenesconsistofmanyclassessimultaneously,whichmakesitcomplextoprocess,sosemanticallysegmentingthemmeansverymuchnotonlyforresearchfieldbutalsoforapplicationfield.Regardingtheproblemsmentionedabove,inthispaper,weproposeamethodthatcombinestheimagesegmentation,whichusesspaceextremepointsastheseedsofwatershedsegmentationandrecognition,whichisbasedontheconditionalrandomfieldmodeltoinferencethesemanticlabelsofeachobjectclass,inaframe.Firstly,accordingtotheframe,aGaussianpyramidofinputimageisbuilt,thenspaceextremepointsoftheimageareextractedtofeedthewatershedsegmentationalgorithmasseeds.Basedonthebasicrulesofimagesegmentation,noisyregionsareremovedandregionssatisfyingthesimilaritymeasurementaremergedtogethertogettheunsupervisedsegmentationresults.Currentresearchresultsshowthatusingappearancefeaturesonlyfailstogetgoodobjectboundaries,soapartfromthebroadlyusedappearance,shapeandcontextfeatures,weaddfivegeometryfeatures,heightabovethecamera,closestdistancetocamerapath,surfaceorientationandtrackdensity,whichcontainrichstructureinformation,toourframework.Morethanthat,wein