Probabilistic Spatio-temporal 2D-Model for Pedestrian Motion Analysis in Monocular Sequences

This paper addresses the problem of probabilistic modelling of human motion by combining several 2D views. This method takes advantage of 3D information avoiding the use of a complex 3D model. Considering that the main disadvantage of 2D models is their restriction to the camera angle, a solution to this limitation is proposed in this paper. A multi-view Gaussian Mixture Model (GMM) is therefore fitted to a feature space made of Shapes and Stick figures manually labelled. Temporal and spatial constraints are considered to build a probabilistic transition matrix. During the fitting, this matrix limits the feature space only to the most probable models from the GMM. Preliminary results have demonstrated the ability of this approach to adequately estimate postures independently of the direction of motion during the sequence.