Download files
Abstract
This thesis offers contributions and innovations to the development of vision-based autonomous flight control systems for small unmanned aerial vehicles operating in cluttered urban environments. Although many optic flow algorithms have been reported, almost none have addressed the critical issue of accuracy and reliability over a wide dynamic range of optic flow. My aim is to rigorously develop improved optic flow sensing to meet realistic mission requirements for autonomous navigation and collision avoidance.
A review of related work enabled development of a new hybrid optic flow algorithm concept combining the best properties of image correlation and interpolation with additional innovations to enhance accuracy, computational speed and reliability. Key analytical work yielded a methodology for determining optic flow dynamic range requirements from system and sensor design parameters and a technique enabling a video sensor to operate as a passive ranging system for closed loop flight control.
Detailed testing led to development of the hybrid image interpolation algorithm (HI2A) using improved correlation search strategies, sparse images to reduce processing loads, a solution tracking loop to bypass the more intensive initial estimation process, a frame look-back method to improve accuracy at low optic flow, a modified interpolation technique to improve robustness and an extensive error checking system for validating outputs.
A realistic simulation system was developed incorporating independent, precision ground truthing to assess algorithm accuracy. Comparison testing of the HI2A against the commonly-used Lucas Kanade algorithm demonstrates major improvement in accuracy over greatly expanded dynamic range.
A reactive flight controller using ranging data from a monocular, forward looking video sensor and rules-based logic was developed and tested in Monte Carlo simulations of a hundred flights. At higher flight speeds than reported in similar tests, collision-free results were obtained in a realistic urban canyon environment. The HI2A algorithm and flight controller software performance on a common PC was up to eight times faster than real-time for outputs of 250 measurements at 50 Hz.
The feasibility of terrain mapping in real-time was demonstrated using 3D ranging data from optic flow in an overflight of the urban simulation environment indicating the potential for its use in path planning approaches to navigation and collision avoidance.