Layer-based video registration

Registration of a mission video sequence with a reference image without any metadata (camera location, viewing angles, and reference DEMs) is still a challenging problem. This paper presents a layer-based approach to registering a video sequence to a reference image of a 3D scene containing multiple layers. First, the robust layers from a mission video sequence are extracted and a layer mosaic is generated for each layer, where the relative transformation parameters between consecutive frames are estimated. Then, we formulate the image-registration problem as a region-partitioning problem, where the overlapping regions between two images are partitioned into supporting and nonsupporting (or outlier) regions, and the corresponding motion parameters are also determined for the supporting regions. In this approach, we first estimate a set of sparse, robust correspondences between the first frame and reference image. Starting from corresponding seed patches, the aligned areas are expanded to the complete overlapping areas for each layer using a graph-cut algorithm with level set, where the first frame is registered to the reference image. Then, using the transformation parameters estimated from the mosaic, we initially align the remaining frames in the video to the reference image. Finally, using the same partitioning framework, the registration is further refined by adjusting the aligned areas and removing outliers. Several examples are demonstrated in the experiments to show that our approach is effective and robust.