融合注意力机制的图神经网络运动动作识别研究  

Research on motor action recognition based on graph neural network integrating attention mechanism

作　　者：张琼[1] ZHANG Qiong(Xi’an Institute of Innovation,Yan’an University,Xi’an 710100,China)

机构地区：[1]延安大学西安创新学院,西安710100

出　　处：《自动化与仪器仪表》2025年第3期196-200,共5页Automation & Instrumentation

基　　金：陕西省体育局体育科研常规课题《关于老年群体体育健身服务需求的研究》(2022040)。

摘　　要：为提高运动动作的精准性,提出一种融合注意力机制的图神经网络运动动作识别方法。该方法以双流自适应时空图神经网络作为基础模型,通过添加边缘边与时空注意力模块,提升模型对动作特征的学习效果,实现了运动动作识别。实验结果表明,该方法相较于基础模型,对运动动作识别的Top-1与Top-5准确率分别提高了3.66%与2.27%,达到了39.48%和62.23%,相较于Deep-LSTM、GA-GCN、ST-GCN等热门运动识别模型更高,在实际体育运动中对不同类别运动动作的平均识别准确率也达到66.67%,具有良好的运动动作识别效果。In order to improve the accuracy of motor action,a new motor action recognition method based on graph neural network is proposed.In this method,the two-flow adaptive spatiotemporal graph neural network is used as the basic model.By adding edge and spatiotemporal attention modules,the learning effect of the model on action features is improved,and the action recognition is realized.The experimental results show that compared with the basic model,the accuracy of Top-1 and Top-5 motion recognition of the proposed method is increased by 3.66% and 2.27% respectively,reaching 39.48% and 62.23%,which is higher than the popular motion recognition models such as Deep-LSTM,GA-GCN and ST-GCN.In actual sports,the average recognition accuracy of different types of movement is 66.67%,which has a good recognition effect of movement.

关 键 词：图神经网络 动作识别 体育训练 时空注意力 运动动作 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]