基于摄像头域内域间合并的无监督行人重识别方法陈利文叶锋黄添强黄丽清翁彬徐超胡杰（福建师范大学计算机与网络安全学院福州350117）（福建省公共服务大数据挖掘与应用工程技术研究中心（福建师范大学）福州350117）（数字福建大数据安全技术研究所（福建师范大学）福州350117）（yefeng@fjnu.edu.cn）AnUnsupervisedPersonRe-IdentificationMethodBasedonIntra-/Inter-CameraMergerChenLiwen,YeFeng,HuangTianqiang,HuangLiqing,WengBin,XuChao,andHuJie（CollegeofComputerandCyberSecurity,FujianNormalUniversity,Fuzhou350117）（FujianProvincialEngineeringResearchCenterofBigDataAnalysisandApplication(FujianNormalUniversity),Fuzhou350117）（DigitalFujianInstituteofBigDataSecurityTechnology(FujianNormalUniversity),Fuzhou350117）AbstractIncriminalinvestigation,intelligentmonitoring,imageretrievalandotherfields,personre-identificationhasalwaysbeenahotandsignificantresearchtopic.Sincemostoftheexistingmethodsrelyonlabeleddatasets,thelackoflabelsmakesunsupervisedpersonre-identificationtechnologymorechallenging.Inordertoovercometheproblemofthelackoflabels,anewframeworktogeneratereliablepseudolabelsisproposedassupervisioninformationforexistingsupervisedpersonre-identificationmodels.Assumingthatimagestakenbythesamecameramainlyvaryinpedestrian’sphysicalappearancesratherthanbackgrounds,thenimagestakenbydifferentcamerasvaryinbackgrounds.Toeliminateeffectsbroughtbythedifferencesinimagebackground,imagesaredividedintoseveraldomainsbycameraSerialnumbersasthefirststep.ThenweconstructundirectedgraphsforeachcamerawithEuclideandistanceofimagepairs,andtherewillbeanedgeonlywhentwoimagesarecloseenough.Theimagesinonemaximumanalogoussetareregardedasthesameperson.Thenwemergemaximumcliquesfromdifferentcamerassimplywiththeircosinedistanceandassignpseudolabels.Theproposedframeworkrunsinanunsupervisedmanner,andtheproposedmethodcanobtainhigheraccuracythantheotherrelatedmethodsonMarket1501andDukeMTMC-ReIDdatasets,whichfurthershowstheeffectivenessoftheproposedmethod.Keywordspersonre-identification；unsupervisedlearning；maxcliquealg...
