Motion-Let Clustering for Skeleton-Based Action Recognition

The representation of 3D skeleton is critical for human action recognition. In this paper, we propose a new approach for skeleton-based action recognition by motion-let clustering. To discover discriminative action features, action sequences are represented by the 3D movements of joints among neighbor frames, called motion-lets. We divide the joints into four groups according to the root-to-end structure, and project the motion-lets to 2D projections. In each 2D projection, the 2D motion-lets of each group are clustered to make primitive representation of actions as histograms. The temporal pyramid is employed to capture the temporal ordering of motion-lets followed by the Random Forest algorithm for classification. The proposed method is validated on four benchmark 3D human action datasets and the experimental results show that the proposed approach outperforms state-of-the-art deep-learning-based methods.