Rapid Self Organizing Maps for Terrain Surface Reconstruction

Kurzfassung

Since their introduction by Kohonen Self-Organizing Maps (SOMs) have been used in various forms for purposes of surface reconstruction. They offer robust and fast approximations of manifold data from unstructured input points while being modestly easy to implement. On the other hand SOMs have certain disadvantages when used in a setup where sparse, reliable and spacial unbounded data occurs. For example, airborne Lidar sensors generate a continuous stream of point data while flying above terrain. We introduce modifications of the SOM's data structure to adapt it to unbound data. Furthermore, we introduce a new variation of the learning rule called rapid learning that is feasible for sparse but rather reliable data. We demonstrate examples where the surroundings of an aircraft can be reconstructed in almost real time.