3DSCAN: Online Ego-Localization and Environment Mapping for Micro Aerial Vehicles
M. Sanfourche, A. Plyer, A. Bernard-Brunel, G. Le Besnerais (Onera)
We present 3DSCAN (3D Scene Characterization for Autonomous Navigation), a software application for state estimation and environment modeling using lowcost 3D sensors, such as a stereorig and RGBD cameras. For state estimation, we describe an original keyframe-based stereoscopic visual odometry technique, which can run at more than 20Hz on a lightweight computer. This so-called ‘efficient Visual Odometry’ (eVO) has been evaluated on several datasets and provides accurate results and limited drift, even for indoor/outdoor trajectories. Environment modeling aggregates instantaneous depthmaps in a volumetric Octomap  representation. Stereoscopic depthmaps are computed by a very fast dense matching algorithm derived from eFolki, an optical flow code implemented on GPU. These developments are combined in the 3DSCAN software, which is successfully demonstrated on our MAV (Micro Aerial Vehicle) system, following indoor, outdoor or mixed trajectories.