语义引导的遮挡行人再识别注意力网络  

Semantic guidance attention network for occluded person re-identification

作　　者：任雪娜 张冬明 包秀国[1,3] 李冰[4] REN Xuena;ZHANG Dongming;BAO Xiuguo;LI Bing(Institute of Information Engineering,Chinese Academy of Sciences,Beijing 100093,China;School of Cyber Security,University of Chinese Academy of Sciences,Beijing 100093,China;National Computer Network Emergency Response Technical Team/Coordination Center of China,Beijing 100029,China;School of Aeronautic Science and Engineering,Beijing University of Aeronautics and Astronautics,Beijing 100191,China)

机构地区：[1]中国科学院信息工程研究所,北京100093 [2]中国科学院大学网络空间安全学院,北京100093 [3]国家计算机网络应急技术处理协调中心,北京100029 [4]北京航空航天大学自动化学院,北京100191

出　　处：《通信学报》2021年第10期106-116,共11页Journal on Communications

基　　金：国家重点研发计划基金资助项目(No.2018YFB0804704);国家自然科学基金资助项目(No.61672495,No.U1736218)。

摘　　要：为了解决遮挡场景下行人再识别的特征不对齐、错误匹配的问题,提出了一种语义引导对齐的注意力网络(SGAN)对齐行人的不同部分。SGAN以行人的语义掩膜作为监督信息,通过全局语义引导和局部语义引导提取行人的全身和局部特征,并根据人体不同部分的可见性动态调整模型训练。在推理阶段,依据注意力模型获得局部区块的可见性,利用共享可见的人体部分的匹配策略自适应地对特征进行相似度的计算。实验结果表明,SGAN能够容忍一定的遮挡,它的准确率不仅在全身数据集上优于大多数先进模型,在2个较大规模的复杂遮挡数据集Occluded-Duke MTMC和P-DukeMTMC-reID上也优于现有的行人再识别方法。To solve the problem of misalignment and mismatch in occluded person Re-ID,SGAN(semantic guided attention network)was proposed.In SGAN,the semantic masks of pedestrians were used as supervision to learn the global and local features through the attention modules,and the training process was dynamically adjusted according to the visibility of local regions.In the inference stage,the part-to-part matching strategy was adopted to adaptively measure visible features based on the feature visibility,which was obtained based on the learned masks from the attention modules.Experimental results show that the average accuracy of SGAN on the holistic datasets is better than most advanced models.Additionally,it is tolerant of occlusions and largely outperforms existing person Re-ID methods on two larger-scale complex occlusion datasets(Occluded-DukeMTMC and P-DukeMTMC-reID).

关 键 词：深度学习 遮挡行人再识别 注意力网络 语义引导 特征对齐 

分 类 号：TN92[电子电信—通信与信息系统]