赵彦如, 牛东杰, 杨蕙萌.基于注意力机制和姿态识别的行人再识别[J].河南理工大学学报（自然科学版）,2023,42(2):120-126.

ZHAO Y R, NIU D J, YANG H M,et al.Person re-identification based on attention mechanism and gesture recognition[J].Journal of Henan Polytechnic University(Natural Science) ,2023,42(2):120-126.

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

doi:10.16186/j.cnki.1673-9787.2021070059

基金项目:国家自然科学基金资助项目（51505133）；河南省科技攻关计划项目（212102210316）；河南理工大学光电传感与智能测控河南省工程实验室开放课题资助项目（HELPSIMC-2020-006）

收稿日期:2021/07/16

修回日期:2021/10/14

出版日期:2023/03/25

Person re-identification based on attention mechanism and gesture recognition

ZHAO Yanru, NIU Dongjie, YANG Huimeng

School of Mechanical and Power Engineering，Henan Polytechnic University，Jiaozuo 454000，Henan，China

Abstract:In the person re-identification of the posture change，occlusion，background and other issues，in order to improve the performance of person re-identification under occlusion，a person re-recognition method based on the combination of attention mechanism and gesture recognition was proposed.The global attention network and the posture recognition network were used to extract the global features，the thermal map of the joint position and the corresponding confidence，respectively.The local features of pedestrians at 13 joint points and fusion of all joints were obtained by calculation.Pedestrian classification training was conducted for global features and 14 local features respectively，and multi-task learning was used to learn multiple losses to jointly supervise network optimization.The Rank-1/mAP values tested on the two large-scale public datasets Market1501 and DukeMTMC-reID reached 85.1%/75.6% and 64.3%/55.3%，respectively.It was shown that the designed method with high recognition accuracy had excellent ability resist to posture changes，occlusion and background.

Key words:deep learning;person re-identification;attention mechanism;gesture recognition;feature fusion;local feature;global feature

 016_2021070059_赵彦如_H.pdf