Human motion recognition method based on dynamic spatial modeling and multi-scale temporal modeling

A new human motion recognition method based on dynamic spatial modeling and multi-scale temporal modeling is proposed. Traditional methods for exploring graph topology often use the same topology for every channel, which is a rigid operation. To solve these problems, a dynamic spatial modeling module is constructed to model an independent spatial topology for each frame skeleton, and a self-attentional diagram convolution module is introduced to enhance the extraction of internal features of the skeleton. The multi-scale temporal convolutional network is constructed by expanding convolution with different expansion rates, and the multi-scale features are extracted. In addition, according to the physical connection between skeletal joints, bone information, joint movement information and bone movement information are extracted from joint information, and multi-level features are input into the network for training and fusion of classification results.