Mixed-state causal modeling for statistical KL-based motion texture tracking

Abstract

We are interested in the modeling and tracking of dynamic or motion textures, which refer to dynamic contents that can be classified as a texture with motion (fire, smoke, crowd of people). Experimentally we observe that they depict motion maps with values of a mixed type: a discrete value at zero (absence of motion) and continuous non-null motion values. We thus introduce a temporal mixed-state Markov model for the characterization of motion textures from which a set of 13 parameters is extracted as the descriptive feature of the dynamic content. Then, a motion texture tracking strategy is proposed using the conditional Kullback?Leibler (KL) divergence between mixed-state probability densities, which allows us to estimate the position using a statistical matching approach.