Vision for vehicle guidance using two road cues

摘要

An autonomous vehicle must be able to find and maintain visual contact with a negotiable path (a road) in front of it. Presents a strategy for road tracking that uses two road cues in order to maintain better visual contract with a road than could be achieved with either cue alone. Also, using two cues to find the road, it is possible to study the ability of a system to recover when contradictory evidence exists concerning the location of the road in front of the vehicle. The two cues used are the road surface shading and its boundaries. The road surface properties captured in an image are modeled by bi-variate polynomial surface patches of up to second order. The road boundaries, and other line-like features of a road, are modeled by line segments fit to image edges. The surface patches and line segments that model the road in one image are used along with knowledge of the vehicle's motion to predict the appearance of the road in a subsequent image. At each image frame the models are updated to take into account new aspects of the road's appearance.