3D Human Motion Prediction Based on Graph Convolution Network and Transformer

Extracting, recognizing and predicting human actions from image information plays an essential part in the fields of human intention understanding, behavior emergency avoidance and automatic driving. In recent years, with deep learing method developing rapidly, the methods of behavior detection and intention understanding for human actions are also glowing with new vitality. In this paper, based on spatial-temporal synchronous graph convolution network and multi-head self-attention mechanism, a new method of human skeleton action recognition and prediction is proposed. By extracting the spatial features of short-term time series at the same time, we can predict the long-term time series actions, and also we have achieved satisfactory experimental results. Our experiment is based on Human3.6M dataset for training and testing. At the end of the paper, we put forward the limitations of the current research and some future research directions.