An Effective 3D Geometric Relational Feature Descriptor for Human Action Recognition

Abstract

This paper presents an effective feature descriptor for recognizing human actions from three-dimension (3D) motion capture video sequences. The proposed feature descriptor is extended from the Boolean features which have been successfully used in computer animation. We first transform 3D coordinates of specified human points, as provided by the motion capture data system, into corresponding 3D points as defined in an articulated 3D human model. We then derive novel 3D geometric relational features, a numeric (continuous-valued) version of the Boolean features, to represent the geometric relations among body points of a pose. Finally, the proposed feature descriptor is applied in human action classification using the hidden Markov model. The simulation results indicate the effectiveness of the proposed feature descriptor as evidenced by the high recognition rate.