Human Action Recognition Based on Multi-Scale Feature Augmented Graph Convolutional Network

Abstract

Nowadays, video has gradually become the mainstream media of communication, and the massive amounts of videos bring challenge to the task of manual review of the videos. So, using computers to understand the videos is of great significance. Among the approaches of automatic action recognition, skeleton-based approach has many advantages, such as strong robustness to light changes, strong action expression ability, small amount of computation time, etc. In this paper, a multi-scale feature augmented graph convolutional network is proposed. It uses the spatial multi-scale GCN module to extract spatial features of different scales, the multi-scale temporal augmentation module to capture temporal features of different scales. To prove the performance of the proposed method, experiments were performed on two public datasets, NTU-RGB+D and The Kinetics-Skeleton. Compared with other advanced action recognition methods, the proposed method can accomplish action recognize effectively, and the recognition accuracy is improved.