基于图卷积的骨架行为识别  

Skeleton Action Recognition Based on Graph Convolution

作　　者：董安 左劼[1] 孙频捷[2] DONG An;ZUO Jie;SUN Jie-pin(College of Computer Science,Sichuan University,Chengdu 610065;Shanghai University of Political Science and Law,Shanghai 200000)

机构地区：[1]四川大学计算机学院,成都610065 [2]上海政法学院,上海200000

出　　处：《现代计算机》2020年第2期48-50,56,共4页Modern Computer

基　　金：国家重点研发计划项目

摘　　要：近年来,随着视频大规模增长,如何高效组织和分析这些海量视频,面临巨大的挑战,而人作为视频中的主体,如何有效分析其行为,成为学术界研究的一个热点。通过利用OpenPose,提取视频中人体的动态骨架序列,采用基于图卷积神经网络的方法,借助在大规模视频动作数据集上训练好的模型进行迁移学习,所提出的方法在自采集的室内监护视频数据集上,取得0.9的识别准确率。In recent years,the large-scale growth of video,how to organize and analyze these massive videos faces enormous challenges,and how to analyze effectively people’s behavior as the main body of video has become a research hotspot in academia.By using OpenPose,extracting the dynamic skeleton sequence of the characters in the video,using the method based on the graph convolutional neural network,and using the model trained on the large-scale video action dataset for transfer learning,on the constructed indoor monitoring video dataset,recogni⁃tion accuracy of 0.9 was obtained.

关 键 词：行为识别 骨架序列 图卷积神经网络 迁移学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]