A new feature descriptor for 3D human action recognition

A novel approach for recognizing human actions using sequences of 3D point clouds of agents over time is presented. It is claimed that some regions that have a long distance to the body center (boundary regions of human body) are very discriminative for understanding human actions. Based on this idea, a novel descriptor based on weighted boundary of 3D point cloud is introduced to recognize the actions. Unlike previously published methods, this descriptor is invariant to scale, translation, and rotation. A dynamic time alignment technique is used as a similarity measure for classification. Experimental results on i3DPost dataset demonstrate the effectiveness of proposed method compared to other existing methods.