Action tree convolutional networks: Skeleton-based human action recognition

Abstract

This paper is mainly about addressing the problem of skeleton-based human activity recognition: ignoring the structure and relationship between skeleton joints and body-parts, the existence of a large amount of useless information in the activity data, and poor generalization ability. In order to solve the shortcomings of existing mainstream methods used for human action recognition, we propose a novel method named Action Tree Convolutional Networks (ATCNs). This method uses a data based auto-designed Action Tree network to dynamically generate a tree of nodes/body-parts and a semantic attention center, profoundly emphasizing the relations and semantics of nodes/body-parts. This method we introduced has a great improvement on the previous algorithm’s neglect of the importance of nodes/body-parts relation, and improves the generalization ability of the algorithm. Through experiments on Kinetics and NTU-RGB+D datasets, our method achieves better performance improvements over other state-of-the-art methods.