A novel 3D human action reconstruction method

Abstract

This paper presents a simple yet effective 3D posture carving reconstruction method for human action recognition. The idea is inspired by the observation that human could carve a 3D object according to a few number of 2D images collected from multiple cameras. First, human action sequences are captured simultaneously by 3 web cameras. Then 3 normalized binary human silhouettes are obtained by a series of preprocessing steps. Finally, the 3D posture in the virtual coordinate is reconstructed according to 2D silhouettes of 3 different views. Experimental results show the effectiveness of the proposed approach.