基于全局和局部令牌变换的行人再识别算法  

Person re-identification algorithm based on global and local token transformer

作　　者：吴婷婷 王苹宇 陈洪刚[2] WU Ting-Ting;WANG Ping-Yu;CHEN Hong-Gang(China Nuclear Power Engineering Company Limited,Beijing 100840,China;College of Electronics and Information Engineering,Sichuan University,Chengdu 610065,China)

机构地区：[1]中国核电工程有限公司,北京100840 [2]四川大学电子信息学院,成都610065

出　　处：《四川大学学报(自然科学版)》2025年第2期399-411,共13页Journal of Sichuan University(Natural Science Edition)

基　　金：国家自然科学基金(62301346);四川省科技计划资助项目(2024NSFSC1424);成都市技术创新研发项目(2024-YF05-00652-SN);四川大学引进人才科研启动经费资助项目(YJ202326);成都市科技成果转化示范项目(2023-YF09-00019-SN);四川大学-中国核动力研究设计院联合创新基金。

摘　　要：行人再识别在智慧安防、智慧城市和智慧交通等领域具有广泛的研究和应用价值,但是就目前的研究和应用需求来看,面向真实场景的行人再识别仍然是一项挑战.针对真实场景下行人目标具有复杂的图像变化,本文提出一种全局和局部令牌变换(GLTT)框架,以学习具有鉴别性和鲁棒性的行人特征.首先,在GLTT框架上引入一种全局令牌变换(GTT)模块,考虑到单个类别令牌难以应对复杂的行人图像变化,该模块利用多个类别令牌从不同语义空间中学习多个全局行人特征,提升行人再识别模型的全局鲁棒性.然后,考虑到行人局部细节包含关键身份信息,本文设计一种局部令牌变换(LTT)模块,通过融合自注意力权重来动态选择具有语义相关性的块令牌,并在所选择的块令牌和类别令牌之间进行信息交互,从而提升行人再识别模型的局部鉴别性.最后,本文提出一种简单有效的类别令牌正则化(CTR)方法,以使每个类别令牌的特征空间不重叠,从而提高多个类别令牌特征的表示能力.实验结果表明,与多种行人再识别方法相比,本文所提出的GLTT框架在Market1501、CUHK03、DukeMTMC和MSMT17数据集上均取得最优的识别效果,验证了此框架具有良好的鉴别性和鲁棒性.The field of person Re-Identification holds significant research and application value in the domains of intelligent security,intelligent city and intelligent transportation.However,considering the current research and application requirements,it is still challenging to implement person Re-Identification models for real-world scenarios.In this work,a Global and Local Token Transformer(GLTT)framework is proposed to effectively capture the complex image variations of person targets,enabling the acquisition of discriminative and robust person features.Firstly,a Global Token Transformer(GTT)module is introduced into the GLTT framework.Considering that a single class token is difficult to deal with complex image variations,the GTT module uses multiple class tokens to learn multiple global features from different semantic spaces and improve the global robustness of person Re-Identification models.Then,since the local details contain crucial person information,a Local Token Transformer(LTT)module is designed to dynamically select semantically relevant patch tokens by fusing self-attention weights.In addition,the LTT module contributes to information interaction between the selected patch tokens and class tokens,and therefore enhances the local discrimination of person re-identification models.Finally,a simple yet effective approach named Class Token Regularization(CTR)method is proposed to enhance the representation capability of multiple class token features by ensuring non-overlapping feature spaces for each class token.The experimental results demonstrate that the proposed GLTT framework achieves superior Re-Identification performance on Market1501,CUHK03,DukeMTMC,and MSMT17 datasets,thereby validating the discrimination and robustness of the proposed framework.

关 键 词：行人再识别 全局令牌变换 局部令牌变换 类别令牌正则化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]