Multi-Scale Feature Extraction Model with Motion Attention for Human Motion Prediction

Abstract

Skeleton-based human motion prediction is an increasingly popular work in computer vision, and it has important guiding roles for many practical applications in our real life, such as video surveillance, human motion synthesis, and human-computer interaction. We observed that in real life, human tends to repeat previous motions, especially periodic motion like walking. Moreover, we can obtain poses at multiple scales by merging multiple joins connected into one join recursively to obtain spatial-temporal features at multiple scales. In this paper, we propose a multi-scale feature extraction model with motion attention for human motion prediction. By fusing features at multiple scales, we can extract features with richer relationships in spatial dimension as well as temporal dimension. Then, the fusion features are fed into the multi-scale motion predictor to obtain the future motion sequence. We conducted experiments on Human 3.6M dataset. The performance is higher than that of the latest method. The experimental results prove the validity of the proposed model.