基于卷积残差学习的行人检测与跟踪.pdf

《基于卷积残差学习的行人检测与跟踪.pdf》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、硕士学位论文基于卷积残差学习的行人检测与跟踪CONVOLUTIONALRESIDUALLEARNINGFORPEDESTRIANDETECTIONANDTRACKINGKOTARGAREMSIMPLICE哈尔滨工业大学2018年6月国内图书分类号：TP309学校代码10213国际图书分类号：681.324密级：公开工学硕士学位论文心脏代谢计算模型的构建及其与电生理机制的耦合实现硕士研究生：KotarGaremSimplice导师：AssociateProf.FengJiang申请学位：工学硕士学科：计算机科学与技术所在单位：计算机科学与技术学院答辩日期：2018年6月授予学位

2、单位：哈尔滨工业大学ClassifiedIndex:TP309U.D.C.:681.324DissertationfortheMaster’sDegreeinComputerinEngineeringCONVOLUTIONALRESIDUALLEARNINGFORPEDESTRIANDETECTIONANDTRACKINGCandidate:KotarGaremSimpliceSupervisor:AssociateProf.FengJiangAcademicDegreeAppliedfor:MasterofEngineeringSpeciality:ComputerScie

3、nceandTechnologyAffiliation:SchoolofComputerScienceandTechnologyDateofDefence:June2018Degree-Conferring-Institution:HarbinInstituteofTechnologyAbstractVisualobjecttrackinghavebeenaninterestingcomputervisionproblematicwithnumerousreal-worldapplications.Bydefinition,itisaprocedureofidentifyin

4、g,locating,anddefiningthedynamicconfigurationofoneormanymoving,possiblydeformableobjectsorpartsofobjectsineachframeofoneorseveralcameras.Theyonlyrequireaminorsetoftrainingsamplesontheinitialframetocreateanappearancemodel.Nevertheless,someotherexistingdiscriminativecorrelationfilterslearnthe

5、filtersdistinctlyfromfeatureextraction,andupdatethesefiltersbymeansofamovingaverageoperationwithanempiricalweight.TheseDCFtrackersbarelyprofitfromtheend-to-endtraining.ThemaindifficultyinVisualobjecttrackingishowtoutilizetheextremelylimitedtrainingdatatodevelopanappearancemodelrobusttoavari

6、etyofchallengesincludingscalevariation,backgroundclutter,partialocclusionsandmotionblur.Thisprojectthereforeattemptstoelucidatetheseissuesthroughthisapproachoverthecurrenthighestaccurategenerativemodel,tracking-by-detectionbasedMDNet[1]andthecorrelation-basedCREST,followedwithaworkontheeffe

7、ctoftheconvolutionalnetworkdepthonitsprecisioninthelarge-scaleimagerecognitionsetting.Atthisstagetheexperimentalresultsonbothstandardbenchmarks,OTB-2013,OTB-2015andVOT2016[2],showimprovementsinspeed,precision,androbustnessonbothtrackers.Keywords:residual