Abstract
In this paper, we tackle the shadow problem in depth for a better foreground segmentation. We propose a novel variant of co-training technique for shadow detection and removal in uncontrolled scenes. This variant works according to a powerful temporal behavior. Setting co-training parameters is based on an extensive experimental study. The proposed co-training variant runs periodically to obtain more generic classifier, thus improving speed and classification accuracy. An experimental study by quantitative, qualitative and comparative evaluations shows that the proposed method can detect shadow robustly and remove the 'cast' part accurately from videos recorded by a static camera and under several constraints.