基于骨架的人体动作识别技术研究进展

Research Progress in Skeleton-Based Human Action Recognition

近年来,随着深度学习技术的发展,已有很多新颖的基于骨架的人体动作识别算法被提出,极大地推动了该领域的发展.对基于骨架的人体动作识别领域的主要数据集和算法进行全面、细致的总结.首先对NTU,Kinet-ics-Skeleton和SYSU 3DHOI等骨架相关的数据集进行回顾;然后将基于骨架的人体动作识别算法归纳为基于监督学习的、基于半监督学习的和基于无监督学习的3大类,并对分属不同类别的算法进行介绍和比较;最后分析和总结得出该领域当前面临过度依赖大数据、大算力和大模型等挑战,并针对性地提出缓解以上挑战的3点未来发展方向:高精度骨架数据集建设、细粒度骨架动作识别和数据有效学习的骨架动作识别.

In recent years,with the development of deep learning technology,many novel skeleton-based human action recognition algorithms have been proposed,which has greatly promoted the development of this field.This paper aims to give a comprehensive and detailed summary of the main datasets and algo-rithms in the skeleton-based human action recognition field.Firstly,the main skeleton-related datasets such as NTU,Kinetics-Skeleton,and SYSU 3DHOI are reviewed.Secondly,the skeleton-based human action recognition algorithms are summarized into three categories,i.e.,supervised learning-based,semi-supervised learning-based,and unsupervised learning-based,the main algorithms of each category are further introduced and compared.Finally,challenges that the field is currently facing,i.e.,over-reliance on big data,large computing power,and large models,are concluded,and three future development directions are proposed to alleviate the above challenges:high-precision skeleton dataset construction,fine-grained skeleton-based action recognition,and skeleton-based action recognition with data-efficient learning.