2023-05-10计算机应用,JournalofComputerApplications2023,43(5):1372-1377ISSN1001-9081CODENJYIIDUhttp://www.joca.cn融合人体全身表观特征的行人头部跟踪模型张广耀1,2,宋纯锋1,2*(1.中国科学院大学人工智能学院,北京100049;2.中国科学院自动化研究所智能感知与计算研究中心,北京100190)(∗通信作者电子邮箱chunfeng.song@nlpr.ia.ac.cn)摘要:现有的行人多目标跟踪模型在密集场景下存在行人无法检出以及帧间关联混淆的问题。为了提高密集场景下行人跟踪的精确率,提出一种融合全身表观特征的行人头部跟踪模型HT-FF(HeadTrackingwithFull-bodyFeatures)。首先,使用行人头部检测器替代全身检测器,提高密集场景下行人的检出率;其次,利用人体姿态估计的信息为引导,获得去噪声的全身表观特征作为跟踪线索,大幅减少多帧之间关联时发生的混淆。HT-FF模型在密集场景下行人跟踪的基准数据集HeadTracking21(HT21)上的MOTA(MultipleObjectTrackingAccuracy)和IDF1(IDF1Score)等多个指标上取得了最优的结果。HT-FF模型能有效缓解密集场景下行人跟踪丢失和混淆的问题,所提出的融合多线索的跟踪模型是行人跟踪任务的新范式。关键词:多目标跟踪;运动模型;动态模型;特征匹配;行人头部跟踪;行人重识别;人体姿态估计;表观特征中图分类号:TP391.4文献标志码:APedestrianheadtrackingmodelbasedonfull-bodyappearancefeaturesZHANGGuangyao1,2,SONGChunfeng1,2*(1.SchoolofArtificialIntelligence,UniversityofChineseAcademyofSciences,Beijing100049,China;2.CenterforResearchonIntelligentPerceptionandComputing,InstituteofAutomation,ChineseAcademyofSciences,Beijing100190,China)Abstract:Theexistingpedestrianmulti-objecttrackingalgorithmshavetheproblemsofundetectablepedestriansandinter-frameassociationconfusionindensescenes.Inordertoimprovetheprecisionofpedestriantrackingindensescenes,aheadtrackingmodelbasedonfull-bodyappearancefeatureswasproposed,namelyHT-FF(HeadTrackingwithFull-bodyFeatures).Firstly,theheaddetectorwasusedtoreplacethefull-bodydetectortoimprovethedetectionrateofpedestriansindensescenes.Secondly,usingtheinformationofhumanpostureestimationasaguide,thenoise-removedfull-bodyappearancefeatureswereobtainedastrackingclues,whichgr...