Adaptive color model for figure-ground segmentation in dynamic environments

Abstract

In this paper we propose a new technique to perform figure-ground segmentation in image sequences of scenarios with varying illumination conditions. Most of the algorithms in the literature that adapt color, assume smooth color changes over time. On the contrary, our technique formulates multiple hypotheses about the next state of the color distribution (modelled with a Mixture of Gaussians -MoG-), and validates them taking into account shape information of the object. The fusion of shape and color is done in a stage denominated 'sample concentration', that we introduce as a final step to the classical CONDENSATION algorithm. The multiple hypotheses generation, allows for more robust adaptions procedures, and the assumption of gradual change of the lighting conditions over time is no longer necessary.