Vision-Based 3D Navigation for an Autonomous Helicopter

We address the 3D navigation problem for an Unmanned Aerial Vehicle (UAV) through urban environments, specifically an autonomous robotic helicopter. In doing so, we investigate techniques which would allow a UAV to perform two kinds of navigation: safe wandering and path planning based on a partially accurate world model, with dynamic replanning as unmodeled obstacles are detected. Typically UAVs operate at high altitudes where obstacle avoidance is not needed. This limits the applications they can be used for. A UAV capable of safe wandering through an urban environment could perform tasks such as exploration, mapping and Urban Search and Rescue. To perform safe wandering, the UAV would need to detect and avoid obstacles in real-time. Having the ability to plan a path through an urban environment would allow a UAV to perform tasks such as package delivery, surveillance and patrolling. In developing these capabilities, we address obstacle avoidance, path planning, replanning, and system integration with a real helicopter. Although obstacle avoidance and path planning are well studied problems for ground-based robots operating in 2D, they are not well studied for integrated 3D problems such as the one proposed in this thesis. We present a novel vision-based obstacle avoidance technique combining optic flow and stereo vision. Through experiments we show that this combined technique produces more reliable safe wandering than either technique alone. Experiments on the real autonomous helicopter show that optic flow can be used to avoid obstacles to the side, and that stereo vision can be used to avoid obstacles to the front. We investigate a number of factors that influence the performance of the optic flow-based centering response, including optimum camera angle, xiv egomotion compensation, and control strategies. Through analytical and empirical investigations, we show the optimum camera angle is 45 degrees. We address the path planning problem by implementing a Probabilistic Roadmap pathplanner, and assume a partially accurate 3D model of the environment is available (implying the need for dynamic replanning). We show 3D path planning in a simulation environment, including dynamic replanning when unmodeled obstacles are detected using stereo vision.