Compact mapping in plane-parallel environments using stereo vision

In this paper we propose a method for transforming a 3D map of the environment, composed by a cloud of millions of points, into a compact representation in terms of basic geometric primitives, 3D planes in this case. These planes, with their texture, yield a very useful representation in robot navigation tasks like localization and motion control. Our method estimates the main planes in the environment (walls, floor and ceiling) using point classification, based on the orientation of their normal and its relative position. Once we have inferred the 3D planes we map their textures using the appearance information of the observations, obtaining a realistic model of the scene.