R3DG features: Relative 3D geometry-based skeletal representations for human action recognition

Recently introduced cost-effective depth sensors coupled with real-time skeleton extraction algorithms have generated a renewed interest in skeleton-based human action recognition. Most of the existing skeleton-based approaches use either the joint locations or the joint angles to represent the human skeleton. In this paper, we introduce and evaluate a new family of skeletal representations for human action recognition, which we refer to as R3DG features. The proposed representations explicitly model the 3D geometric relationships between various body parts using rigid body transformations, i.e., rotations and translations in 3D space. Using the proposed skeletal representations, human actions are modeled as curves in R3DG feature spaces. Finally, we perform action recognition by classifying these curves using a combination of dynamic time warping, Fourier temporal pyramid representation and support vector machines. Experimental results on five benchmark action datasets show that the proposed representations perform better than many existing skeletal representations. The proposed approach also outperforms various state-of-the-art skeleton-based human action recognition approaches.