Appearance-based Data Association for 3D and Multisensory SLAM in Structured Environment

Simultaneous Localization and Mapping (SLAM) has been studied a lot since 20 years in the robotics community: a SLAM method copes with the incremental construction of a geometrical model from the successive fusion of observations acquired from moving sensors. When sensors are embedded on a mobile robot, robot localization and environment mapping must be performed in the same function. Many methods aim to build sparse representations, made of 2D segments (from 2D laser data for indoor applications) or 3D points (from a single camera or stereovision): this paper discusses some issues related to 3D and dense SLAM, i.e. the incremental construction of a textured surfacic model from the fusion of sensory data acquired from a 3D laser range finder coupled with a camera: the fusion of multisensory data in a unified 3D representation makes more robust the data association between features perceived from different viewpoints. First an heterogeneous model, made of a stochastic map and of appearance-based information, is described, before focusing on the data association function; our approach is validated by experimental results obtained from data acquired on our Jido robot.