骨骼双流注意力增强图卷积人体姿态识别  

Bone Dual-Stream Attention Enhancement Graph Convolving Human Posture Recognition

作　　者：陈斌 樊飞燕 陆天易 Chen Bin;Fan Feiyan;Lu Tianyi(Information Construction Management Division,Nanjing Normal University,Nanjing 210023,China)

机构地区：[1]南京师范大学信息化建设管理处,江苏南京210023

出　　处：《南京师范大学学报（工程技术版）》2024年第4期57-67,共11页Journal of Nanjing Normal University(Engineering and Technology Edition)

基　　金：江苏省现代教育技术研究2023年度智慧校园专项课题(2023-R-107311).

摘　　要：为解决骨骼关键点分类算法中运动时间线中运动关联信息的价值分析缺乏,以及骨骼节点关联性和依赖关系信息含义丢失问题,提出了一种骨骼双流注意力增强图卷积人体姿态识别模型.以提取骨骼特征节点为基础,构建骨骼关节点之间空域连接矩阵和运动时间线时域信息矩阵,在此基础上进行双流骨骼节点信息处理.利用通道注意力机制对上下文处理的优势,构解关键节点间依赖关系以及全局骨骼运动含义,构建邻域节点加权的双域骨骼拓扑.在Kinetics和NTU RGB+D两个数据集上的对比验证显示,该模型在不同数据集上均有较好的执行效果.与领域内较具代表性的主流方法的横向比对显示,该模型在选定的9种行为姿态的识别精度上均优于其他模型.该方法在人体姿态识别上体现了较优的识别率及稳定性,并佐证了时空双域骨骼特征信息的挖掘价值.In order to solve the lack of value analysis of motion correlation information in the loss of meaning of skeletal nodes and dependency information,the paper proposes a model of bone dual-stream attention enhancement graph convolving human posture recognition.The airspace connection matrix and time domain information matrix between bone joints are constructed on the basis of extracting bone feature nodes.With this basis,dual-flow bone node information processing is performed.Taking advantage of the channel attention mechanism for context processing,decturing key node dependencies and global bone motion implications,a two-domain bone topology weighted by neighborhood nodes is constructed.The comparative validation on two datasets Kinetics and NTU RGB+D shows that the model performs better on different datasets.Horizontal comparison with the more representative mainstream methods in the field is shown,the model outperforms the other models in the recognition accuracy of the nine selected behavioral poses.This method reflects the better recognition rate and stability in human posture recognition,and proves the mining value of spatial-temporal dual-domain bone feature information.

关 键 词：姿态识别 时空双域 注意力机制 图卷积 骨骼特征 运动信息表示 

分 类 号：TP39[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]