Adaptive Foreground/Background Segmentation Using Multiview Silhouette Fusion

We present a novel approach for adaptive foreground/background segmentation in non-static environments using multiview silhouette fusion. Our focus is on coping with moving objects in the background and influences of lighting conditions. It is shown, that by integrating 3d scene information, background motion can be compensated to achieve a better segmentation and a less error prone 3d reconstruction of the foreground. The proposed algorithm is based on a closed loop idea of segmentation and 3d reconstruction in form of a low level vision feedback system. The functionality of our approach is evaluated on two different data sets in this paper and the benefits of our algorithm are finally shown based on a quantitative error analysis.