Journal of Frontiers of Computer Science and Technology ›› 2021, Vol. 15 ›› Issue (4): 641-657.DOI: 10.3778/j.issn.1673-9418.2008088

• Surveys and Frontiers • Previous Articles     Next Articles

Survey on Two-Dimensional Human Pose Estimation of Deep Learning

ZHOU Yan, LIU Ziqin, ZENG Fanzhi, ZHOU Yuexia, CHEN Jiawen, LUO Yue   

  1. Department of Computer Science, Foshan University, Foshan, Guangdong 528000, China
  • Online:2021-04-01 Published:2021-04-02

深度学习的二维人体姿态估计综述

周燕刘紫琴曾凡智周月霞陈嘉文罗粤   

  1. 佛山科学技术学院 计算机系,广东 佛山 528000

Abstract:

In recent years, as a hot spot in the field of computer vision, human pose estimation has broad application prospects in video surveillance, human-computer interaction, and intelligent campus. With the rapid development of neural networks, the use of deep learning methods for 2D human pose estimation, compared with traditional methods that require manual setting of features, can more fully extract image information, and obtain more robust characteristics, so deep learning based methods have become the mainstream of 2D human pose estimation algorithm research. However, deep learning is still developing, there are still problems such as large training scale, and researchers mainly start with designing networks and training methods to improve the human pose estimation algorithm. Fristly, the two-dimensional human pose estimation is divided into two categories: single person and multiple persons. Secondly, single-person pose estimation is divided into coordinate regression and heat map detection according to different ground truth types, and multi-person pose estimation is divided into two-step and single-step method according to different algorithm steps, to summarize and classify the advanced algorithms in recent years, and their advantages and disadvantages as well as scope of application are analyzed. Then, the international standard datasets and the corresponding evaluation indices are introduced, and some classical algorithms are compared experimentally. Finally, the current research problems and the future development trends are summarized.

Key words: human pose estimation, deep learning, neural networks, joints detection

摘要:

近年来人体姿态估计作为计算机视觉领域的热点,在视频监控、人机交互、智慧校园等领域具有广泛的应用前景。随着神经网络的快速发展,采用深度学习方法进行二维人体姿态估计,相较于传统需要人工设定特征的方法,更能充分地提取图像信息,获取更具有鲁棒性的特征,因此基于深度学习的方法已成为二维人体姿态估计算法研究的主流方向。然而,深度学习尚在发展中,仍存在训练规模大等问题,研究者们主要从设计网络以及训练方式入手对人体姿态估计算法进行改进。首先,将二维人体姿态估计分为单人与多人两大类进行论述;根据真值类型不同将单人姿态估计分为基于坐标回归与基于热图检测两类,根据算法步骤不同将多人姿态估计分为二步法与一步法两类,对近年来先进的算法进行总结分类介绍,并分析它们的优缺点以及适用范围;然后,介绍了相关的国际标准数据集以及相应的评价指标,并对几种经典算法进行实验数据对比;最后,对当前研究所存在的问题以及未来发展趋势进行了总结概述。

关键词: 人体姿态估计, 深度学习, 神经网络, 关节点检测