Integrated Stereovision for an Autonomous Ground Vehicle Competing in the Darpa Grand Challenge

Sammanfattning: The DARPA Grand Challenge (DGC) 2005 was a competition, in form of a desert race for autonomous ground vehicles, arranged by the U.S. Defense Advanced Research Project Agency (DARPA). The purpose was to encourage research and development of related technology. The objective of the race was to track a distance of 131.6 miles in less than 10 hours without any human interaction. Only public GPS signals and terrain sensors were allowed for navigation and obstacle detection. One of the teams competing in the DGC was Team Caltech from California Institute of Technology, consisting primarily of undergraduate students. The vehicle representing Team Caltech was a 2005 Ford E-350 van, named Alice. Alice had been modified for off-road driving and equipped with multiple sensors, computers and actuators. One type of terrain sensors used on Alice was stereovision. Two camera pairs were used for short and long range obstacle detection. This master thesis concerns development, testing and integration of stereovision sensors during the final four months leading to the race. To begin with, the stereovision system on Alice was not ready to use and had not undergone any testing. The work described in this thesis enabled operation of stereovision. It further improved its capability such that it increased the overall performance of Alice. Reliability was demonstrated through multiple desert field tests. Obstacle avoidance and navigation using only stereovision was successfully demonstrated. The completed work includes design and implementation of algorithms to improve camera focus and exposure control, increase processing speed and remove noise. Also hardware and software parameters were configured to achieve best possible operation. Alice managed to qualify to the race as one of the top ten vehicles. However she was only able to complete about 8 miles before running over a concrete barrier and out of the course, as a result of hardware failures and state estimation errors.