Skeleton-based Action Recognition Using Two-stream Graph Convolutional Network with Pose Refinement

With the development of science and technology, graph convolutional network has made great progress in improving the accuracy of action recognition. However, there still exists some deficiencies in current methods. Firstly, the human skeleton point coordinates entering into the network are barely refined, which may cause large error. Secondly, the second-order information(the length and direction of bones), which can reflect action characteristics discriminatively, is rarely used. To solve the above issues, a two stream graph convolutional network with pose refinement for skeleton based action recognition is proposed. Besides, we use an adaptive block to to help improve the accuracy. We test our method on Kinetics dataset and the experiment show it can get better results than some recent methods, which plays a positive role in future research.