Range image segmentation and surface parameter extraction for a mobile robot

This paper presents a methodology for a concise representation of the 3-D world model for a mobile robot, using range data. The process starts with the segmentation of the scene into ''objects'' that are given a unique label, based on principles of range continuity. Then the external surface of each object is partitioned into homogeneous surface patches. Contours of surface patches in 3-D space are identified by estimating the normal and curvature associated with each pixel. The resulting surface patches are then classified as planar, convex or concave. Since the world model uses a volumetric representation for the 3-D environment, planar surfaces are represented by thin volumetric polyhedra. Spherical and clindrical surfaces are extracted and represented by appropriate volumetric primitives. All other surfaces are represented using the boolean union of spherical volumes (as described in a separate paper by the same authors). The result is a general, concise representation of the external 3-D world, which allows for efficient and robust 3-D object recognition. 20 refs., 14 figs.