计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (9): 51-66.DOI: 10.3778/j.issn.1002-8331.2110-0300

• 热点与综述 • 上一篇    下一篇

基于深度学习的行人重识别综述

杨永胜,邓淼磊,李磊,张德贤   

  1. 1.河南工业大学 信息科学与工程学院,郑州 450001
    2.河南省粮食信息处理国际联合实验室,郑州 450001
  • 出版日期:2022-05-01 发布日期:2022-05-01

Overview of Pedestrian Re-Identification Based on Deep Learning

YANG Yongsheng, DENG Miaolei, LI Lei, ZHANG Dexian   

  1. 1.College of Information Science and Engineering, Henan University of Technology, Zhengzhou 450001, China
    2.Henan International Joint Laboratory of Grain Information Processing, Zhengzhou 450001, China
  • Online:2022-05-01 Published:2022-05-01

摘要: 行人重识别主要研究在不同摄像机拍摄的图形中检索目标行人的任务,是计算机视觉领域一个极具挑战性的研究课题。传统依赖手工特征的行人重识别方法性能低且鲁棒性差,不能适应数据爆炸增长的信息时代。近年来,随着大规模行人数据集的出现和深度学习的迅速发展,行人重识别研究取得了许多突出成果。梳理了性能接近饱和的有监督学习研究方法,并探讨近几年研究热度较高的弱监督学习、跨模态数据和端到端的行人重识别现状;对不同类型行人重识别方法比较分析,列举了常用数据集,并将部分经典算法在Market-1501、DukeMTMC-ReID等数据集上进行性能比较;对行人重识别的未来研究方向进行了展望。

关键词: 行人重识别, 有监督学习, 弱监督学习, 跨模态, 端到端

Abstract: Person re-identification(ReID) aims to retrieve target pedestrians across multiple non-overlapping cameras, which is a challenging research topic in the field of computer vision. The traditional person re-identification methods relying on manual features have low performance and poor robustness, and can not adapt to the information age of data explosion. In recent years, with the emergence of large-scale pedestrian data sets and the rapid development of deep learning, many outstanding results have been achieved in pedestrian re-identification research. This paper first sorts out the research methods of supervised learning with performance close to saturation, and discusses the current situation of weakly supervised learning, cross-modal data and end-to-end pedestrian re-recognition, which has been studied in recent years. After that, comparison and analysis of different types of person re-identification methods and the commonly used data sets are listed. The performance of some classic algorithms is compared on data sets such as Market-1501 and DukeMTMC-ReID. Finally, the future research direction of pedestrian re-recognition is prospected.

Key words: person re-identification, supervised learning, weakly supervised learning, cross-modality, end-to-end