基于双流交互学习的长时空换装行人重识别  

Long-Term Spatio-Temporal Clothing Changing Person ReID Based on Interacting Dual-Stream Learning

作　　者：钟铭恩[1] 邓智颖 袁彬淦 谭佳威 杨凯博 ZHONG Mingen;DENG Zhiying;YUAN Bingan;TAN Jiawei;YANG Kaibo(Fujian New Energy Vehicle and Safety Technology Research Institute,Xiamen University of Technology,Xiamen,Fujian 361024,China;School of Aerospace Engineering,Xiamen University,Xiamen,Fujian 361102,China)

机构地区：[1]厦门理工学院福建省新能源汽车与安全技术研究院,福建厦门361024 [2]厦门大学航空航天学院,福建厦门361102

出　　处：《北京理工大学学报》2024年第8期838-849,共12页Transactions of Beijing Institute of Technology

基　　金：福建省自然科学基金资助项目(2023J011439,2019J01859)。

摘　　要：近年来深度行人重识别已取得快速进展,但长时空换装问题仍颇具挑战.为此构建一种双流交互学习算法模型(interactive dual-stream learning,IDSL):基于现有公开数据集生成无服装辅助模态图像,构建主辅双流分支网络来分别学习原始图像和无服装模态图像的细粒度特征;设计一种多尺度特征级联融合器(multi-scale feature cascade fusion,MSFCF),对细粒度特征进行重组并引进交叉注意力机制来实现全局语义和局部细节间的联合建模,提升模型鲁棒性;提出一种具有软惩罚机制的颈部网络(soft penalty batch normalization neck network,SoftBNNeck)来更好地区分度量学习和分类学习,使模型训练更稳定和可控;最后定义了双流一致性约束损失(dual-stream consistency constraint loss,DCCLoss)并探索了多损失联合训练策略,以更好地衡量换装行人身份的概率分布差异,提升重识别准确度.实验表明,在复杂换装行人公开数据集LTCC和Celeb-reID上,Rank-1/mAP分别达到73.8%/47.9%和66.7%/22.6%,领先于同类研究算法.In order to challenge difficulty of the long-term cross-domain re-identification in deep person re-iden-tification(ReID)with clothing changes,an interactive dual-stream learning(IDSL)algorithm model was pro-posed.Firstly,the clothing-agnostic modal images were generated from existing public datasets,and a dual-stream network with main and auxiliary branches was constructed to learn fine-grained features from the origin-al images and modal images,respectively.Then a multi-scale feature cascade fusion module(MSFCF)was de-signed to reorganize the fine-grained features and introduce a cross-attention mechanism in jointly modeling for global semantics and local details,enhancing model robustness.And then a soft penalty batch normalization neck network(SoftBNNeck)was proposed to distinguish metric learning with classification learning,making the training model stable and controllable.Finally,a dual-stream consistency constraint loss(DCCLoss)was defined,and multi-loss joint training strategies were explored to better measure the probability distribution difference of cross-domain person identities,improving re-identification accuracy.Experiment results show that based on the challenging LTCC and Celeb-reID cross-domain person re-identification datasets,IDSL can achieve Rank-1/mAP scores of 73.8%/47.9%and 66.7%/22.6%respectively,outperforming prior methods in this field.

关 键 词：深度学习 换装行人重识别 TRANSFORMER 双流网络 细粒度特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] X956[自动化与计算机技术—计算机科学与技术]