Combined shape analysis of human poses and motion units for action segmentation and recognition

Recognizing human actions or analyzing human behaviors from 3D videos is an important problem currently investigated in many research domains. The high complexity of human motions and the variability of gesture combinations make this task challenging. Local (over time) analysis of a sequence is often necessary in order to have a more accurate and thorough understanding of what the human is doing. In this paper, we propose a method based on the combination of pose-based and segment-based approaches in order to segment an action sequence into motion units (MUs). We jointly analyze the shape of the human pose and the shape of its motion using a shape analysis framework that represents and compares shapes in a Riemannian manifold. On one hand, this allows us to detect periodic MUs and thus perform action segmentation. On another hand, we can remove repetitions of gestures in order to handle with failure cases for the task of action recognition. Experiments are performed on three representative datasets for the task of action segmentation and action recognition. Competitive results with state-of-the-art methods are obtained in both the tasks.