Color vision for the detection of unstructured road and intersections

Abstract

My research addresses the problem of navigating a real robot vehicle on unstructured roads, which have no lane markings, may have degraded surfaces and edges, and may be partially obscured by strong shadows. These conditions cause many road following systems to fail. We have built two separate systems, SCARF and UNSCARF, that are based on standard pattern recognition techniques to successfully navigate on a variety of unstructured roads. SCARF determines a likelihood that each pixel in the image belongs to the road by comparing the color of each pixel to sampled road and off-road colors. UNSCARF uses a standard clustering technique to group pixels with similar features. Both SCARF and UNSCARF then use a technique based on matching models of road shape to locate the roads in the image. This method is more robust in noisy conditions than other road interpretation techniques. This method is reliable enough for SCARF to be the first system to detect intersections without using a priori knowledge of the intersection shape and location predicted from a map. SCARF and UNSCARF have been integrated into several navigation systems which have successfully driven a test vehicle on a variety of test sites and in many types of weather conditions.