In this paper a new technique to perform figure-ground segmentation in image sequences of scenarios with varying illumination conditions is proposed. The set of color points of both the target and background are modelled with Mixture of Gaussians (MoG), which optimum number is automatically initialized. Based on the 'Linear Discriminant Analysis' (LDA) a new colorspace that maximizes the foreground/background class separability is presented. Moreover, there is no need to assume gradual change of the viewing conditions over time, because the method works with multiple hypotheses about the next state of the color distribution (some considering small changes and other more abrupt variations). The hypothesis that generates the best object segmentation and the shape information in the previous iteration are fused to accurately detect the object boundary, in a stage denominated 'sample concentration', introduced as a final step to the classical CONDENSATION algorithm.