Skeleton based human action recognition with relative position encoding

Abstract

Aiming at the fact that the current graph convolution operation based on skeleton graph is limited in local adjacent nodes, or the overall relative position information of skeleton is omitted, an enhancement method of joint point position information based on relative position encoding of skeleton is proposed. The proposed method takes the central joint point of the human trunk as the root node, and all joint nodes form a tree structure according to the natural connection of the body, and the code of each joint node inherits the code of its parent node and also includes its own number in the sibling node. In addition, considering that the number of channels in the graph-based convolutional network model is generally larger, and the channel information itself has a strong correlation, the channel information frequency division and recombination operation is proposed to reflect the difference of information in different frequency bands in the channel. Experiments show that the proposed method has a certain effect on improving the effect of the embedded model.