基于时空图小波神经网络的手部动作识别方法

根据车间人员操作监控的需要，文中研究了一种基于深度学习的新方法——时空图小波神经网络（ST-GWNN）。该算法对图小波卷积进行参数化，以降低每层图卷积层的参数复杂度，并采用一阶切比雪夫多项式逼近图小波卷积；分离多项式阶数 K 与邻接节点阶数之间的关系，固定多项式阶数，通过调整超参数 s 来改变邻域范围，从而识别更多复杂的手部动作。实验结果表明，文中提出的ST-GWNN在动作识别中的识别率优于目前常用的时空图卷积神经网络，并且能够充分利用动作的时空关联性。

Hand Movements Recognition Method Based on Spatio-temporal Graph Wavelet Convolutional Neural Networks

According to the needs of monitoring the operation of workers, a new deep learning model named as ST-GWNN (spatio-temporal graph wavelet convolutional neural network) is presented in this paper. The algorithm parameterizes the graph wavelet convolution to reduce the parameter complexity of each graph convolution layer, and uses the first-order Chebyshev polynomial to approximate the graph wavelet convolution. The relation between the polynomial order K and the order of adjacent nodes is separated. Fix the polynomial order, and change the neighborhood range by adjusting the hyperparameter s , so as to recognize more complex hand movements. Experimental results show that the hand movements recognition accuracy of the proposed ST-GWNN is better than that of the commonly used spatio-temporal graph convolutional neural network, and the proposed ST-GWNN can make full use of the spatio-temporal correlation of hand movements.