National Science Foundation
Accurate background modeling is fundamentally important to motion-based segmentation, object tracking, and video surveillance. Models must discriminate between coherent foreground motion and periodic, random, or small pixel variations typically found in complex outdoor scenes.
We are developing robust methods that are capable of detecting coherent foreground regions in dynamic outdoor scenes. Our approach involves a multi-resolution match filter framework that models the locally changing spatial image structure at many scales. The correlation values of these filters are combined to robustly discriminate foreground regions from regions that conform to the background model.