面向复杂场景的RGBT目标跟踪方法研究.pdf
Abstract
RGBTtargettrackingaimstoeffectivelyfusevisibleandthermalinfraredvideosequences
together.TherearetwomainreasonswhyRGBTtargettrackingcanachieveall-weatherefficient
monitoring.Ontheonehand,RGBTtargettrackingcaneffectivelyusethermalinfraredinformation
toprovidestronginformationcompensationfortheapparentcharacteristicsofvisiblelighttargets
underpoorlightingconditions.Ontheotherhand,visiblelightinformationcanassistinsolvingthe
thermalcrossproblemfacedinthermalinfraredimageprocessing.Althoughtherearemanyresearch
resultsinrelatedfields,theystillfacechallengessuchasdynamicocclusionandsimilarbackground.
Withtherapiddevelopmentofartificialintelligenceandbigdataprocessingtechnology,deep
learninghasdevelopedinthedirectionofcross-taskandmulti-modal.Thisdevelopmenttrendcan
providesceneinformationforsolvingtheinherentdifficultiesinRGBTtargettracking.Inorderto
effectivelyutilizemulti-tasksceneinformation,thispaperdeeplystudiesRGBTobjecttracking
methodforcomplexscenesfromthreelevels:feature,updatestrategyandattention.Theresearch
innovationsofthispaperareasfollows:
(1)Featurelevel:AimingatthelimitationthattraditionalRGBTtargettrackingmethodsonly
focusonthefeaturerepresentationlearningofthetargettobetracked,atargettrackingmethodbased
onsceneconsistencyisproposed.Thedesignintentionoftheproposedmethodistofindthat
strengtheningtheconsistencyofglobalreasoningofdifferentmodalitiesishelpfultoimprovethe
robustnessoftargetfeaturesandsolvetheproblemofincompletemodalinformationcausedby
complexbackgrounds.Basedonthis,undertheframework