第42卷第2期2023年3月Vol.42No.2Mar.2023JOURNALOFHENANPOLYTECHNICUNIVERSITY(NATURALSCIENCE)河南理工大学学报(自然科学版)基于注意力机制和姿态识别的行人再识别赵彦如,牛东杰,杨蕙萌(河南理工大学机械与动力工程学院,河南焦作454000)摘要:在解决行人再识别技术中的姿态变化、遮挡、背景等问题时,为了提高遮挡下的行人再识别性能,提出一种基于注意力机制和姿态识别的行人再识别方法。采用全局注意网络和姿态识别网络分别提取行人图像的全局特征、关节点位置热力图和对应的置信度,通过计算得到行人13个关节点和融合所有关节点的局部特征,对全局特征和14个局部特征分别进行行人分类训练,利用多任务学习多个损失共同监督网络的优化。测试时,将关键点特征和全局特征融合后,计算行人的距离排序。在Market1501和DukeMTMC-reID数据集上测试的Rank-1/mAP指标分别达到了85.1%/75.6%和64.3%/55.3%。结果表明,所设计方法具备抗姿态变化、遮挡和背景的能力,同时具有较高的识别能力和识别精度。关键词:深度学习;行人再识别;注意力机制;姿态识别;特征融合;局部特征;全局特征中图分类号:TP391文献标志码:A文章编号:1673-9787(2023)2-120-7Personre-identificationbasedonattentionmechanismandgesturerecognitionZHAOYanru,NIUDongjie,YANGHuimeng(SchoolofMechanicalandPowerEngineering,HenanPolytechnicUniversity,Jiaozuo454000,Henan,China)Abstract:Inthepersonre-identificationoftheposturechange,occlusion,backgroundandotherissues,inor⁃dertoimprovetheperformanceofpersonre-identificationunderocclusion,apersonre-recognitionmethodbasedonthecombinationofattentionmechanismandgesturerecognitionwasproposed.Theglobalattentionnetworkandtheposturerecognitionnetworkwereusedtoextracttheglobalfeatures,thethermalmapofthejointpositionandthecorrespondingconfidence,respectively.Thelocalfeaturesofpedestriansat13jointpointsandfusionofalljointswereobtainedbycalculation.Pedestrianclassificationtrainingwasconductedforglobalfeaturesand14localfeaturesrespectively,andmulti-tasklearningwasusedtolearnmultiplelossestojointlysupervisenetworkoptimization.TheRank-1/mAPvaluestestedonthetwolarge-scalepublicdatasetsMarket1501andDukeMTMC-reIDreached85.1%/75.6%and6...