6DOF entropy minimization SLAM for stereo-based wearable devices

In this paper, we propose and validate a novel approach to solve the Simultaneous Localization and Mapping (SLAM), focused on its application with wearable devices. In order to do so, we use a stereo vision camera as the unique sensor that provides semi-dense information of the environment (appearance and range data). A first approximation of the trajectory is given by an egomotion algorithm, that exploits the information of the stereo observations in order to estimate the action between each pair of consecutive observations (visual odometry). The algorithm provides a locally but not globally consistent approximation because it is only based on local information. In order to obtain a globally consistent map, which is the key topic of this paper, we propose an Information Theory based approach that rectifies the map obtained by the egomotion step by performing successive refinements over the trajectory using global information. The key idea is that the best aligned map is the one with the minimum entropy. In order to ensure the scalability of the algorithm, we propose a dynamic map compression strategy that bounds the complexity of the problem and attenuates both memory and computing time requirements. In the experimental section, we show the results of the algorithm in several situations: structured/unstructured environments, indoor/outdoor scenarios, cyclic/acyclic trajectories, etc. performed with a wearable stereo device that we have built to carry out these experiments.

Research highlights
► Fast and robust 6DOF SLAM for Wearable Devices.
► Real-time egomotion with structural refinement.
► Low-error long-range trajectories mapping in both indoor and outdoor environments.
► Information theory based approach to the SLAM problem.
► Information fusion of range and appearance (stereo-based) for map building.