A 2D and 3D registration framework for remote-sensing data

Time series of elevation models have been successfully used to study crustal deformations induced by earthquakes, to monitor landslides and glaciers, and to track dune fields. In all these studies, accurate registration of the data has been a crucial step, which can be seen as finding the transformation that maps each element of a first acquisition to elements of a second acquisition. We answer the registration problem by solving a global energy minimization problem, composed of a matching criterion and a regularization criterion. Both criteria can be easily adapted depending on the context, and we globally minimize this energy with an iterative and multi-scale framework based on fast graph solvers. Contrary to heuristics-based methods such as ICP, we demonstrate that our algorithm can find a solution close to the global optimal solution. In addition, we detail how our method can cope with large dataset through a tiling approach based on a Lagrangian scheme. We demonstrate the versatility of the method for both 2D and 3D applications: (1) Using 2D stereo images in an urban context we show accurate reconstruction of building heights, (2) Using 3D LiDAR point clouds above the White Sands Dunes in New Mexico, we track the 3D motion of the dune field. Glacier monitoring and damage assessment applications are also currently being investigated.