Vision-based path-planning in unstructured environments

Autonomous driving in unstructured environments has attracted an unprecedented level of attention when the DARPA announced the Grand Challenge Competitions in 2004 and 2005. Autonomous driving involves (at least) three major subtasks: perception of the environment, path planning and subsequent vehicle control. Whereas the latter has proven a solved problem, the first two constituted, apart from hardware failures, the most prominent source of errors in both Grand Challenges. This paper presents a system for real-time feature detection and subsequent path planning based on multiple stereoscopic and monoscopic vision cues. The algorithm is, in principle, suitable for arbitrary environments as the features are not tailored to a particular application. A slightly modified version of the system described here has been successfully used in the Qualifications and the Final Race of the Grand Challenge 2005 within the Desert Buckeyes' autonomous vehicle