计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (20): 1-29.DOI: 10.3778/j.issn.1002-8331.2404-0143

• 热点与综述 • 上一篇    下一篇

骨架人体行为识别研究回顾、现状及展望

边存灵,吕伟刚,冯伟   

  1. 1.中国海洋大学 基础教学中心,山东 青岛 266100
    2.天津大学 智能与计算学部,天津 300350
  • 出版日期:2024-10-15 发布日期:2024-10-15

Skeleton-Based Human Action Recognition:History,Status and Prospects

BIAN Cunling, LYU Weigang, FENG Wei   

  1. 1.Teaching Center of Fundamental Courses, Ocean University of China, Qingdao, Shandong 266100, China
    2.College of Intelligence and Computing, Tianjin University, Tianjin 300350, China
  • Online:2024-10-15 Published:2024-10-15

摘要: 人体行为识别在视频监控、人机交互、医疗看护、体育赛事分析等领域具备重要的应用前景。近年来,随着传感器技术和人体姿态估计算法的迅猛发展,基于骨架的人体行为识别受到研究者越来越多的重视。相较于传统的视频图像数据,骨架数据以行为人为中心,具有高度抽象的运动信息和低数据维度等特点,为行为信息建模提供了新的视角。以骨架人体行为识别为研究对象,对相关工作进行了全面系统的回顾和分析。通过文献计量分析法对已发表的相关文献进行了梳理,系统总结了基于骨架的行为识别的发展脉络。在此基础上,分别回顾了基于手工特征的传统识别方法和基于深度学习的识别方法,重点介绍了基于卷积神经网络、循环神经网络、图卷积神经网络以及Transformer方法的基本原理、改进策略和代表性工作,并简要论述了网络模型学习算法的研究现状。总结了基于运动捕捉系统、Kinect相机和RGB图像的三类公开数据集,并详细探讨了它们的特点和应用。最后,结合国内外研究现状及思考分析,梳理了基于骨架的人体行为识别中的关键难题与挑战,并展望了未来的发展方向,旨在为研究人员建立一个较完整的领域研究视图,为相关领域的工作提供参考和借鉴。

关键词: 骨架行为识别, 文献计量分析, 时空特征表征, 深度学习, 神经网络

Abstract: Human action recognition has significant application prospects in fields such as video surveillance, human-computer interaction, medical care, and sports event analysis. In recent years, with the rapid development of sensor technology and human pose estimation algorithms, skeleton-based human action recognition has gained increasing attention from researchers. Compared to traditional video or image data, skeleton data have the characteristics of being centered on the human subject, highly abstract motion information, and low data dimensions, providing a new perspective for modeling behavior information. This paper focuses on skeleton-based human action recognition and provides a comprehensive systematic review and analysis of relevant work. Firstly, through a literature citation analysis, it systematically summarizes the development trajectory of skeleton-based action recognition. Based on this, the paper reviews traditional recognition methods based on manual features and deep learning-based methods, focusing on the basic principles, improvement strategies, and representative works of convolutional neural networks, recurrent neural networks, graph convolutional neural networks, and Transformer methods, and briefly discusses the research status of network model learning algorithms. Secondly, it summarizes three types of publicly available datasets based on motion capture systems, Kinect camera, and RGB images, and discusses their characteristics and applications in detail. Finally, combined with the current research status and thinking analysis at home and abroad, the paper summarizes the key challenges and difficulties of skeleton-based human action recognition, and looks forward to future development directions, aiming to establish a comprehensive domain research perspective for researchers and provide a reference and inspiration for work in related fields.

Key words: skeleton-based action recognition, bibliometric analysis, spatial-temporal information representation, deep learning, neural network