A generic methodology for partitioning unorganised 3D point clouds for robotic vision

Range image segmentation has many applications in computer vision areas such as computer graphics and robotic vision. A generic methodology for 3D point set analysis in which planar structures play an important role is defined. It consists mainly of a specific K-means algorithm which is able to process different shapes in cluster. At the same time, within geometric and topologic considerations, a set of application-driven heuristics is designed. This helps to find out the right number of structures in point sets in order to give a good visualization and representation of a large scale environment without a priori models. Our aim is to propose a simple and generic frame for 3D scene understanding. Tests were realised on different types of environment data: natural and man-made. This research project has been realized with EADS (French Air Space Society).