Manifold regularized particle filter for articulated human motion tracking

Abstract

In this paper, a fully Bayesian approach to articulated human motion tracking from video sequences is presented. First, a filtering procedure with a low-dimensional manifold is derived. Next, we propose a general framework for approximating this filtering procedure based on the particle filter technique. The low-dimensional manifold can be treated as a regularizer which restricts the space of all possible distributions to the space of distributions concentrated around the manifold. We refer to our method as Manifold Regularized Particle Filter. The proposed approach is evaluated using real-life benchmark dataset HumanEva.