Adaptive Region-Based Video Registration

Video registration without meta data (camera location, viewing angles, and reference DEMs) is still a challenging problem. With the aim of handling this kind of problem, this paper presents an adaptive region expansion approach to propagate the alignment process from high confidence areas (reliable salient features) to low confidence areas and to simultaneously remove outlier regions. Hence, we re-cast the image registration problem as a partitioning problem such that we determine the optimal supporting regions and their corresponding motion parameters for the registration. First, we determine sparse robust correspondences between mission and reference images by using our wide baseline algorithm. Next, starting from the seed regions, the aligned areas are expanded to the whole overlapping areas using the graph cut algorithm, which is controlled by the level set representation of the previous region shape. Consequently, a robust video registration is achieved if the scene can be represented by one homography. Furthermore, we extend this approach to multi-homography video registration problem for 3D scenes, which cannot be directly solved by the current alignment methods. Using our motion layer extraction algorithm, the mission video first is segmented into several layers, then each layer is respectively aligned to the reference image by employing the region expansion algorithm. Several examples are demonstrated in the experiments to show that our approach is effective and robust.