基于城市信息单元和差异注意力的多层行人重识别技术  

Multi-level Person Re-identification based on Urban Information Unit and Diff Attention Scheme

作　　者：朱利[1] 林欣 徐亦飞 刘真[2] 马英[3] ZHU Li;LIN Xin;XU Yifei;LIU Zhen;MA Ying(School of Software Engineering,Xi’an Jiaotong University,Xi’an 710049,China;School of Computer and Information Technology,Beijing Jiaotong University,Beijing 100091,China;State Information Center,Beijing 100038,China)

机构地区：[1]西安交通大学电信学部软件学院,西安710049 [2]北京交通大学计算机科学与信息学院,北京100091 [3]国家信息中心,北京100038

出　　处：《集成技术》2023年第1期91-104,共14页Journal of Integration Technology

基　　金：国家重点科研项目(2019YFB2102500)。

摘　　要：在现实的智慧城市安全场景中,传统的行人重识别方法已经难以满足复杂多样的识别任务要求。为实现多层次的行人重识别,该文提出将行人重识别技术与多层次的城市信息单元深度融合。在行人重识别任务中,现有的模型和注意力只关注鲁棒特征的学习,而该文基于特征向量差异,提出了差异注意力模块,以增强深度特征的判别力。结合差异注意力模块,该文开发了与多种骨干模型适配的差异注意力框架。此外,该文还提出了联合训练和单独训练两种训练策略。与其他行人重识别方法相比,差异注意力框架和训练策略在Market-1501、CUHK03和MSMT17数据集上均取得了更优的性能。The traditional person re-identification methods are difficult to independently cope with the complex and diverse recognition tasks in the security scenario of smart city in practice.In order to meet the needs of multi-level person re-identification,the deep integration of person re-identification and multi-level urban information units is proposed.Existing models and attentions for person re-identification tasks only focus on learning the robust features while neglecting the difference between features of pairs.Diff attention module is proposed to guide the network to learn a more discriminative attention map based on the difference of feature vectors.Taking the diff attention module,diff attention framework which matches many backbone models is developed.Two training strategies:joint training and separate training are proposed.Compared with other person re-identification methods,these framework and strategies have achieved excellent performance on Market-1501,CUHK03,and MSMT17 datasets.

关 键 词：行人重识别 城市信息单元 差异注意力 距离函数 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]