计算机技术与发展2025，Vol.35Issue(8)：84-92,9.DOI:10.20165/j.cnki.ISSN1673-629X.2025.0062
基于时序解耦图卷积的行为识别方法
A Temporal Disentanglement Graph Convolutional Network for Action Recognition
摘要
Abstract
Work-related injury prevention is one of the key factors to ensure the safety of employees,reduce economic losses,and promote the stable development of enterprises and society.Most work-related accidents are caused by improper actions.Therefore,intelligent analysis and identification of personnel action is an important measure to prevent work-related accidents.However,existing action recognition methods employ contrasting sample pairs at a single spatiotemporal granularity for unsupervised training,which lacks effective temporal awareness capabilities.Considering that an action originates from the movement changes of body parts on multiple picture frames,decoupling temporal relationships can provide more comprehensive and complementary action clues.Therefore,we introduce a Temporal Disentanglement Graph Convolutional Network(TD-GCN)to improve the modeling ability of action recognition.Specifically,a temporal downsampling module is designed to granularly divide the video sequence and retain the temporal multi-scale features.Then,attention modules are used to conduct global perception modeling of the relevance of temporal clues from different granu-larities.Furthermore,in order to high-light the physical structures embedded in the human body,we introduce a novel spatiotemporal grouping masking strategy that enhances specific spatiotemporal patterns of actions.Experiments conducted on NTU-60 and NTU-120 show that the proposed TD-GCN is superior to several mainstream methods.The developed approach can accelerate the intelligentization,digitization and institutionalization of work injury prevention.关键词
工伤预防/行为识别/时序解耦/图卷积网络/时空分组掩码Key words
work injury prevention/action recognition/temporal disentanglement/graph convolution network/spatiotemporal grouping mask分类
信息技术与安全科学引用本文复制引用
张彬,庄叶,徐天阳,宋晓宁,范方英..基于时序解耦图卷积的行为识别方法[J].计算机技术与发展,2025,35(8):84-92,9.基金项目
国家自然科学基金项目(62106089,62020106012,U1836218) （62106089,62020106012,U1836218）
