基于特征融合的遮挡行人重识别方法

Occluded Person Re-Identification Method Based on Fused Feature

行人重识别是一项解决跨场景跨摄像头下的行人识别问题的技术。当障碍物遮挡行人的某些部位时,人物结构的完整性会被破坏,影响识别效果或难于识别。针对行人重识别的遮挡问题,提出了一种新的遮挡行人重识别方法,引入关系感知全局注意力机制,融合全局特征和局部特征来进行行人重识别。识别方法的主要组成部分包括局部分支、全局分支和特征融合分支等。局部分支通过遍历局部特征来增强鲁棒性;全局分支通过利用关系感知全局注意力来挖掘行人显著性的信息;特征融合分支通过自适应地生成特征权重,利用加权求和的方式来融合全局特征和局部特征。实验验证了所提方法的有效性,并且有效提升了行人重识别的效果。

Person re-identification aims to achieve pedestrian retrieval across different times,places,or cameras.When a person is partially occluded,it might lead to wrong retrieval results if a model does not differentiate the obstruction region and the person region.To solve the occluded person re-identification problem,a new occluded person re-identification method is proposed.The proposed method uses a relation-aware global attention mechanism and fuses global and local features.The main components of the proposed method include local branch,global branch,and feature fusion branch.The local branch enhances robustness by traversing local features.The global branch uses relation-aware global attention to mine pedestrian saliency information.The feature fusion branch adaptively generates feature weights and fuses global feature and local features in a weighted summation way.Experimental results show that the proposed method is effective.