基于相对骨骼点特征和时序自适应感受野的动作识别方法  

Action Recognition Based on Relative Skeleton Point Features and Temporal Adaptive Receptive Field

作　　者：胡昊 史天运 宋永红[3] 余淮 HU Hao;SHI Tian-yun;SONG Yong-hong;YU Huai(Postgraduate Department, China Academy of Railway Sciences, Beijing 100081, China;China Academy of Railway Sciences Corporation Limited, Beijing 100081, China;School of Software, Xi'an Jiaotong University, Xi'an 710049, China;Signal & Communication Research Institute, China Academy of Railway Sciences Corporation Limited, Beijing 100081, China)

机构地区：[1]中国铁道科学研究院研究生部,北京100081 [2]中国铁道科学研究院集团有限公司,北京100081 [3]西安交通大学软件学院,西安710049 [4]中国铁道科学研究院集团有限公司通信信号研究所,北京100081

出　　处：《导航定位与授时》2022年第3期132-139,共8页Navigation Positioning and Timing

基　　金：中国国家铁路集团有限公司系统性重大课题(P2020T002)。

摘　　要：针对长时间动作识别难以充分利用时空域信息的问题,提出了基于相对骨骼点特征和时序自适应感受野的动作识别方法。首先,该方法在特征获取部分增加了相对骨骼点特征,以满足节点多样性和互补性要求,将其分别输入到空域图卷积网络,获得空间中相邻关节聚合的局部特征。然后,设计了一个时序自适应感受野网络,以获取在时域中关节变化的局部特征,并且增加了网络对不同持续时长动作的适应性。最后,经过决策级融合模块,计算类别概率,得到分类结果。仿真结果表明,基于NTU RGB+D和Kinetics-skeleton两大基准数据集,对比多种主流方法,均取得了更高的识别准确率,分别为96.2%与60.1%。该方法可以较好地提取不同动作的区别性时间特征,提高了动作时空特征的判别能力。Aiming at the problem that it is difficult to make full use of temporal and spatial domain information for long-term action recognition,an action recognition method based on relative skeleton point features and temporal adaptive receptive field is proposed.Firstly,this method adds relative skeleton point features in the feature acquisition part to meet the node diversity and complementarity requirements,and respectively inputs them into the spatial graph convolution network to obtain the local features of the adjacent joint aggregation in space.Then,a time-series adaptive receptive field network is designed to obtain the local features of joint changes in the time domain,and to increase the adaptability of the network to actions of different durations.Finally,through the decision-level fusion module,the category probability is calculated to obtain the classification result.The simulation results show that based on the two benchmark datasets of NTU RGB+D and Kinetics-skeleton,compared with other mainstream methods,this method has achieved higher recognition accuracies,which are 96.2%and 60.1%,respectively.The method can better extract the discriminative temporal features of different actions,and improve the discrimination ability of spatiotemporal action features.

关 键 词：动作识别 时序特征提取 图卷积网络 相对骨骼点特征 时序自适应 

分 类 号：U495[交通运输工程—交通运输规划与管理] V249[交通运输工程—道路与铁道工程]