Design and development of stereoscopic obstacle detection systems for UAV

Abstract

Recently a lot of attention has been given to enabling autonomous capabilities on a broad range of vehicles. Among them Unmanned Aerial Vehicles (UAVs), and in particular small quadrotors, must comply to payload-related constraints which poses difficult challenges on the choice of the sensors used to perceive the world. This thesis introduces an obstacle detection and mapping system designed for an UAV based on stereoscopic vision. A small consumer drone, equipped with cheap and lightweight stereo cameras, is used as development robotic platform for this system, for directly demonstrating its applicability. The proposed algorithm is designed to exploit visual information collected from several pair of cameras in order to detect obstacles around the aircraft. Each pair of cameras collect stereoscopic images which are used to recover depth information about the surroundings. Depth quantities incoming from each stereo camera are then fused together and used to build a single representation of the world. This environment description, completely tri-dimensional, is suited for defining any type of environment, because no assumptions have been made on its geometry. The final result is provided as an occupancy grid map, hence conveys information about each portion of the space being occupied or free. Such representation can be used as such, for visualization purposes in an exploration mission, or as main input for subsequent modules in a completely autonomous system. Inverse sensor models for both pinhole and spherical camera are presented and applied during the occupancy grid mapping procedure. A ground truth building approach is presented and used to asses the quality of the obstacles detection and mapping algorithm. The proposed ground truth generation, consisting in exploiting high-accuracy data to build a supposedly true representation of the environment, is tested on a freely available dataset. Results are obtained by comparison between the proposed obstacle detection system and the constructed ground truth.