计算机科学与探索 ›› 2024, Vol. 18 ›› Issue (8): 2014-2033.DOI: 10.3778/j.issn.1673-9418.2311049

• 前沿·综述 • 上一篇    下一篇

无监督学习步态识别综述

陈福仕,沈尧,周池春,丁锰,李居昊,赵东越,雷永升,潘亦伦   

  1. 1. 中国人民公安大学 侦查学院,北京 100038
    2. 大理大学 工程学院,云南 大理 671003
    3. 大理大学 云南省教育厅天空地一体化智能与大数据应用工程研究中心,云南 大理 671003
    4. 中国人民公安大学 公共安全行为实验室,北京 100038
  • 出版日期:2024-08-01 发布日期:2024-07-29

Review of Unsupervised Learning Gait Recognition

CHEN Fushi, SHEN Yao, ZHOU Chichun, DING Meng, LI Juhao, ZHAO Dongyue, LEI Yongsheng, PAN Yilun   

  1. 1. College of Investigation, People??s Public Security University of China, Beijing 100038, China
    2. School of Engineering, Dali University, Dali, Yunnan 671003, China
    3. Air-Space-Ground Integrated Intelligence and Big Data Application Engineering Research Center of Yunnan Provincial Department of Education, Dali University, Dali, Yunnan 671003, China
    4. Public Security Behavioral Science Lab, People??s Public Security University of China, Beijing 100038, China
  • Online:2024-08-01 Published:2024-07-29

摘要: 在光学技术高速发展的现代,步态特征因非接触、非侵入、难伪造、远距离采集等优势受到了学界的广泛关注。目前步态识别算法主要为依赖标签数据的有监督学习方法,庞大的标签标注量在实际应用中面临多重挑战。无监督学习不需要标注就能完成对数据内在特征的自动分析,更贴合实际应用的需求。为了全面认识无监督学习步态识别发展现状及趋势,对领域相关工作进行了梳理。介绍了步态识别常用数据集、通用制作方式以及主流评价指标。从基于GAN的步态识别方法、基于聚类的步态识别方法、基于无监督域适应的步态识别方法和其他方法四个方向详细介绍了目前基于无监督学习的步态识别相关研究思路;选取了CASIA-B、OU-MVLP和OU-ISIR LP三个典型数据集,对主要无监督算法性能进行综合对比;对各方向研究侧重点进行总结讨论,针对存在的交叉研究情况进行评论综述,为未来研究提供借鉴思路。研究分析了无监督步态识别算法目前面临的挑战,并以此展望步态领域未来的发展方向。

关键词: 步态识别, 数字图像处理, 神经网络, 无监督学习, 机器学习, 生物特征识别

Abstract: In the era of fast-paced development of optical technologies, gait analysis has become increasingly significant due to its non-contact, non-invasive nature, resistance to impersonation, and suitability for long-distance data capture. Current gait recognition algorithms mainly use supervised learning, which requires extensive labeled data, and faces practical challenges. Unsupervised learning, which can automatically extract intrinsic features of data without labels, aligns better with real-world needs. This paper reviews the development and trends of unsupervised learning in gait recognition by collating relevant research work. Initially, it outlines commonly used gait datasets, their standard creation methods, and mainstream evaluation metrics. It then delves into the current research on unsupervised learning for gait recognition, detailing approaches from four perspectives: GAN-based methods, clustering-based methods, unsupervised domain adaptation techniques, and other approaches. The performance of major unsupervised algorithms is compared on three typical datasets: CASIA-B, OU-MVLP and OU-ISIR LP. This paper also summarizes the research focus of each direction, and comments on the existence of cross-cutting researches, so as to provide ideas for future research. Lastly, it analyzes the challenges faced by unsupervised gait recognition algorithms and forecasts potential future development in the gait recognition field.

Key words: gait recognition, digital image processing, neural network, unsupervised learning, machine learning, bio-metric identification