Estimation and Tracking of Local Road Geometry for Assisting Reliable Urban Driving

October 16, 2013, 5PM

두산인문관(8) 1 연강홀


For safe urban driving, keeping a car within a road-lane boundary is a critical prerequisite. It requires human and robotic drivers to recognize the boundary of a road-lane and the vehicles location with respect to the boundary of a road-lane that the vehicle happens to be driving in. To provide such a perception capability, we present a new computer vision system that analyzes a stream of perspective images to produce information about a vehicles relative location to a road-lanes boundary, and information about the detecting of lane-crossing and lane-changing maneuvers. To assist the vehicles lateral localization, our algorithm also estimates the host road-lanes geometry, including the number of road lanes and their widths. The local road geometry estimated by frame-by-frame may be inconsistent over frames due to variations in the image features. To handle such inconsistent estimations, we implement a Bayes filter to smooth out, over time, the estimated road geometry. Tests on inter-city highway showed that our system provides stable and reliable performance in terms of computing lateral distances and detecting lane-crossing and lane-changing maneuvers.