Augmented Graph Attention with Temporal Gradation and Reorganization for Human Motion Prediction

Shaobo Zhang, Sheng Liu, Fei Gao, Shengyong Chen

Abstract: Human motion prediction has progressed leap-forward by propulsion of many prior excellent work since its great significance for the promotion of various artificial intelligence applications. How to extract historical spatio-temporal feature better to reduce the discontinuity and long-term error accumulation of prediction motion are still the main challenge of current literature. In this work, a novel augmented graph attention with temporal gradation and reorganization method is proposed, which combines channel attention with graph attention and temporal convolution to be a integrated block for the first time on human motion prediction. The block learns ‘what’, ‘where’ and ‘when’ while capturing the spatial structure of human skeleton in the channel-spatial axes and temporal information in the sequential dimension, respectively. Furthermore, the mechanism of temporal gradation and reorganization can retain complicated and high dynamic temporal information effectively without selection of convolutional kernel size. Our experiments on Human3.6M datasets show that the proposed network performs higher prediction accuracy compared with state-of-the-art methods.