姿态引导与特征增强结合的遮挡行人重识别  

Occluded person re-identification with pose-guided and feature augmentation

作　　者：杜浩宇 苟刚[1,2] DU Hao-yu;GOU Gang(State Key Laboratory of Public Big Data,Guizhou University,Guiyang 550025,China;College of Computer Science and Technology,Guizhou University,Guiyang 550025,China)

机构地区：[1]贵州大学公共大数据国家重点实验室,贵州贵阳550025 [2]贵州大学计算机科学与技术学院,贵州贵阳550025

出　　处：《计算机工程与设计》2024年第6期1843-1849,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(62162010);贵州省科技支撑计划基金项目(黔科合支撑[2022]一般267)。

摘　　要：为解决现有遮挡行人重识别方法只注重于引入外部信息而忽略特征增强的问题,提出一种姿态引导与特征增强结合的遮挡行人重识别方法。将小步幅的滑动窗口引入VIT(Vision-Transformer),使网络获取局部的细微特征,将特征经过遮挡消除模块,消除遮挡带来的干扰;在模型中融入人体姿态估计网络,辅助模型解决遮挡带来的语义信息缺失问题;通过CBN模块提高模型的学习能力,使模型学习到更多高级语义信息。在遮挡行人重识别主流数据集Occluded-DukeMTMC上达到69.8%的Rank-1准确率以及63.2%的mAP,优于现有的其它方法,在整体行人重识别数据集上也取得了具有竞争力的结果。To address the issue that existing occluded person re-identification methods only focus on incorporating external information and ignore feature augmentation,a method was proposed that combining pose guidance and feature augmentation for occluded person re-identification.A small sliding window was introduced into the Vision-Transformer(VIT)to enable the network to acquire fine-grained local features,and the feature was processed through an occlusion elimination module to eliminate interference caused by occlusion.A human pose estimation network was integrated into the model to assist in solving the problem of semantic information loss caused by occlusion.The learning ability of model was improved through a CBN module to enable itself to learn more advanced semantic information.This approach achieves a Rank-1 accuracy of 69.8%and mAP of 63.2%on the mainstream Occluded-DukeMTMC person re-identification dataset,outperforming existing methods.Competitive results are also obtained on the holistic person re-identification dataset.

关 键 词：遮挡行人重识别 姿态引导 特征增强 Vision-Transformer模型 小步幅滑动窗口 CBN模块 遮挡消除模块 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]