Learning Topological Representation of 3D Skeleton Dynamics with Persistent Homology for Human Activity Recognition

The human skeleton is essential in human-computer interaction applications as a typical representation of human activity. Investigating the nonlinear dynamics of the skeleton information could provide an essential understanding of the human locomotion system. This work presents a topological nonlinear analysis approach for the skeleton-based dynamical system in learning the patterns of human activity recognition (HAR). Traditional approaches to dynamics modeling include linear and nonlinear methods with specific characterization capabilities but also drawbacks. With the persistent homology analysis of the point clouds generated via phase space reconstruction, the nonlinear dynamics’ topological descriptors are used to distinguish the activities, which is a novel approach we investigate in this article. As validated with five skeleton-based HAR datasets of MSR Actions 3D, MSR 3D Actions Pair, MSR Daily Activity 3D, UTD-MHAD, and G3D, the topological descriptors of nonlinear dynamics are shown to be excellent features for activity recognition. The proposed topological approach is a promising scheme for video inference in human activity analysis applications.