行人重识别研究综述

Review of Person Re-identification

行人重识别的目的是在多个不重叠的摄像头之间检索特定的行人.对目前有代表性的基于深度学习的行人重识别算法进行归纳和总结,综述不同类型的行人重识别算法的结构和特点.首先,介绍行人重识别的概念;其次,根据行人重识别算法的特点,概述基于监督学习和弱监督学习的行人重识别算法,并对特征表示学习和深度度量学习2种基于监督学习的行人重识别算法进行详细讨论;然后,介绍这一领域的经典数据集,对有代表性的算法在这些数据集上的表现进行对比分析;最后,展望行人重识别领域的发展方向.

The goal of person re-identification(ReID)is to retrieve a specific person between multiple non-overlapping cameras.Representative deep-learning-based person ReID methods were reviewed and the architectures and features of different person re-identification methods were discussed in this paper.First,the definition of person ReID was introduced.Second,supervised learning methods and weakly supervised learning methods were reviewed,respectively.In particular,The feature representation learning and metric learning of the supervised learning methods were discussed in detail.Third,the mainstream datasets in different tasks were analyzed,and the state-of-the-art methods were evaluated and compared.Finally,important yet under-developed research directions were briefed.