Spatially adaptive illumination modeling for background subtraction

Background subtraction is important for many vision applications. Existing techniques can adapt to gradual changes in illumination but fail to cope with sudden changes often seen in indoor environment. In this paper, we propose a novel background subtraction technique that models the change of illumination as a regression function of spatial image coordinates. Such spatial dependency is significant when light sources are close to or within the scene. The regression function is learned from highly probable background regions and applied to the rest of the background models to compensate for the illumination change. While a single regression function is adequate for a smooth Lambertian surface, multiple regression functions are needed to handle depth discontinuities, shadows, and non-Lambertian surfaces. The change of illumination is first segmented and different regression functions are applied to different segments. Experimental results comparing our techniques with other schemes show better foreground segmentation during illumination change.