Data-driven human model estimation for realtime motion capture

Abstract

In this paper, we present a practicable method to estimate individual 3D human model in a low cost multi-view realtime 3D human motion capture system. The key idea is: using human geometric model database and human motion database to establish geometric priors and pose prior model; when given the geometric prior, pose prior and a standard template geometry model, the individual human body model and its embedded skeleton can be estimated from the 3D point cloud captured from multiple depth cameras. Because of the introduction of the global prior model of body pose and shapes into a unified nonlinear optimization problem, the accuracy of geometric model estimation is significantly improved. The experiments on the synthesized data set with noise or without noise and the real data set captured from multiple depth cameras show that the estimation results of our method are more reasonable and accurate than the classical methods, and our method is better noise-immunity. The proposed new individual 3D geometric model estimation method is suitable for online realtime human motion tracking system.