Factored Interval Particle Filtering for Gait Analysis

Commercial gait analysis systems rely on wearable sensors. The goal of this study is to develop a low cost marker less human motion capture tool. Our method is based on the estimation of 3d movements using video streams and the projection of a 3d human body model. Dynamic parameters only depend on human body movement constraints. No trained gait model is used which makes this approach generic. The 3d model is characterized by the angular positions of its articulations. The kinematic chain structure allows to factor the state vector representing the configuration of the model. We use a dynamic Bayesian network and a modified particle filtering algorithm to estimate the most likely state configuration given an observation sequence. The modified algorithm takes advantage of the factorization of the state vector for efficiently weighting and resampling the particles.