Vision-based guidance and control of a hovering vehicle in unknown environments

Abstract

This thesis presents a methodology, architecture, hardware implementation, and results of a system capable of controlling and guiding a hovering vehicle in unknown environments, emphasizing cluttered indoor spaces. Six-axis inertial data and a low-resolution onboard camera yield sufficient information for image processing, Kalman filtering, and novel mapping algorithms to generate a, high-performance estimate of vehicle motion, as well as an accurate three-dimensional map of the environment. This combination of mapping and localization enables a quadrotor vehicle to autonomously navigate cluttered, unknown environments safely. Communication limitations are considered, and a hybrid control architecture is presented to demonstrate the feasibility of combining separated proactive offboard and reactive onboard planners simultaneously, including a detailed presentation of a novel reactive obstacle avoidance algorithm and preliminary results integrating the MIT Darpa Urban Challenge planner for high-level control. The RAVEN testbed is successfully employed as a prototyping facility for rapid development of these algorithms using emulated inertial data and offboard processing as a precursor to embedded development. An analysis of computational demand and a comparison of the emulated inertial system to an embedded sensor package demonstrates the feasibility of porting the onboard algorithms to an embedded autopilot. Finally, flight results using only the single camera and emulated inertial data for closed-loop trajectory following, environment mapping, and obstacle avoidance are presented and discussed.