文献类型：期刊文献

中文题名：基于轻量级图卷积的人体骨架动作识别方法

英文题名：Human Skeleton Action Recognition Method Based on Lightweight Graph Convolution

作者：孙琪翔[1];何宁[2];张聪聪[1];刘圣杰[1]

第一作者：孙琪翔

机构：[1]北京联合大学北京市信息服务工程重点实验室,北京100101;[2]北京联合大学智慧城市学院,北京100101

第一机构：北京联合大学北京市信息服务工程重点实验室

年份：2022

卷号：48

期号：5

起止页码：306-313

中文期刊名：计算机工程

外文期刊名：Computer Engineering

收录：CSTPCD;;北大核心:【北大核心2020】；CSCD:【CSCD_E2021_2022】；

基金：国家自然科学基金(61872042,61572077);北京市教委科技计划重点项目(KZ201911417048);北京市教委科技计划面上项目(KM202111417009);北京联合大学人才强校优选计划(BPHR2020AZ01,BPHR2020EZ01);北京联合大学科研项目(ZK50202001);北京联合大学研究生科研创新项目(YZ2020K001)。

语种：中文

中文关键词：人体骨架动作识别;数据融合;图卷积;非局部网络模块;Ghost网络

外文关键词：human skeleton action recognition;data fusion;graph convolution;non-local network module;Ghost network

摘要：视频中的人体动作识别在计算机视觉领域得到广泛关注,基于人体骨架的动作识别方法可以明确地表现人体动作,因此已逐渐成为该领域的重要研究方向之一。针对多数主流人体动作识别方法网络参数量大、计算复杂度高等问题,设计一种融合多流数据的轻量级图卷积网络,并将其应用于人体骨架动作识别任务。在数据预处理阶段,利用多流数据融合方法对4种特征数据流进行融合,通过一次训练就可得到最优结果,从而降低网络参数量。设计基于图卷积网络的非局部网络模块,以捕获图像的全局信息从而提高动作识别准确率。在此基础上,设计空间Ghost图卷积模块和时间Ghost图卷积模块,从网络结构上进一步降低网络参数量。在动作识别数据集NTU60 RGB+D和NTU120 RGB+D上进行实验,结果表明,与近年主流动作识别方法ST-GCN、2s AS-GCN、2s AGCN等相比,基于该轻量级图卷积网络的人体骨架动作识别方法在保持较低网络参数量的情况下能够取得较高的识别准确率。
Human action recognition in video has garnered extensive attention in the field of computer vision.The action recognition method based on human skeleton can clearly represent human motion;therefore,it has gradually become one of the most important research directions in the abovementioned field.To solve the issue of numerous network parameters and high computational complexity in most mainstream human action recognition methods,a lightweight graph convolution network integrating multistream data is designed and applied to human skeleton action recognition.In the data preprocessing stage,the multistream data fusion method is used to fuse four characteristic data streams.Optimal results can be obtained via one round of training;as such,the number of network parameters required is reduced.A non-local network module based on graph convolution network is designed to capture the global information of an image to improve the accuracy of action recognition.Subsequently,a space Ghost graph convolution module and a time Ghost graph convolution module are designed to further reduce the number of network parameters from the network structure.Experiments are performed on action recognition datasets NTU60 RGB+D and NTU120 RGB+D.Results show that compared with recent mainstream action recognition methods STGCN,2s AS-GCN,and 2s AGCN,the human skeleton action recognition method based on the lightweight graph convolution network can achieve higher recognition accuracy while maintaining a lower number of network parameters.