基于图卷积网络的人体骨架行为识别方法综述

A REVIEW OF HUMAN SKELETON ACTION RECOGNITION METHODS BASED ON GRAPH CONVOLUTIONAL NETWORK

基于骨架数据的人体行为识别已成为计算机视觉领域最热门和最重要的研究课题之一。相较于其他数据类型,人体骨架数据不受光照、背景、视角变化的影响,使得该类行为识别方法具有更强的鲁棒性。此外,骨架数据是以拓扑图结构的形式存在,而图卷积是一种基于图结构的深度学习方法,能够高效地对人体骨架数据的特征进行提取和分类。因此,基于图卷积的方法已经成为处理骨架数据的主流。针对基于图卷积的行为识别方法的前沿性,对其进行全面和系统的总结和分析具有十分重要的意义。本文主要对基于图卷积方法行为识别技术的最新进展进行全面的综述,对相关方法进行分类与总结,并对基准数据集进行详细研究,最后讨论了未来的研究方向和趋势。

Skeleton-based action recognition has emerged as one of the most prominent and crucial research areas in computer vision.The use of human skeleton data in action recognition provides a higher degree of robustness to changes in light,background,and perspective,compared to other data modalities.Human skeleton data exists in the form of a topological graph structure,and graph convolution,which is a deep learning method based on graph structures,is capable of extracting and classifying the features of human skeleton data in an efficient manner.Therefore,graph convolution-based methods have become the mainstream approach for processing skeleton data.It is essential to comprehensively and systematically summarize and analyze the action recognition methods based on graph convolution.Accordingly,this paper provides a comprehensive review of the latest advancements in action recognition technology based on graph convolution methods.The relevant methods are classified and summarized,and the benchmark dataset is studied in detail.Finally,the future research direction and trend are discussed.