Computer vision based markerless human motion tracking has gained popularity in various potential application domains including automatic visual surveillance, security, human computer interaction, virtual reality and medical applications. In computer vision tracking, articulated human body is a very challenging issue because of unknown motion types and high dimensionality. The low-dimension approaches have been effective for overcoming the high-dimensionality problem of tracking the various motions. In this paper, we present a manifold motion model learning in low-dimensional subspace using charting, a nonlinear dimension reduction technique which identify and extract the manifold action from the high-dimensional space. We choose the kernel regressor with Relevance Vector Machine (RVM) to construct the interface between action joint configuration and image space (e.g., Silhouette). The proposed framework allows the identification of the learning phase forward and backward mapping. For tracking of all generative components of the framework we proposed the use of Quantum-inspired particle swarm optimization algorithm to handle local minima problem also for providing global optimization results in search space. © 2012 The Authors.