Path Planning for Autonomous Driving Based on Stereoscopic and Monoscopic Vision Cues

This paper presents a system for real-time feature detection and subsequent path planning for autonomous driving. Special focus lies on visual features for autonomous navigation: stereoscopic and monoscopic cues are employed to distinguish between trafficable road and obstacles of any kind; temporal and stereoscopic point correspondences are used to determine the ego motion of the vehicle. A probabilistic framework for path planning is formulated that models the a priori knowledge about a path and the likelihoods of the visual features. Our path planning algorithm employs a Bayes filter approach that allows recursive integration of new measurements. A slightly modified version of the system was successfully used at the qualifications and final race of the DARPA Grand Challenge 2005 within the Desert Buckeyes' autonomous vehicle. However, the algorithm is, in principle, suitable for arbitrary environments and not limited to unstructured terrain.