Progress In Road Intersection Detection For Autonomous Vehicle Navigation

Abstract

The autonomous land vehicle (ALV) has successfully navigated simple non-branching roads. To traverse more complicated road-ways, the ALV must be able to detect road intersections. We present two methods of road intersection detection. The first recognizes an image road intersection by matching image road boundary data with a predefined intersection model. The predefined model is the intersection configuration derived from an a priori map database. Road boundaries are extracted from the image and processed to produce a set of linear segments. A heuristic method produces a set of combinations of these segments that correspond to segments in the model. The best match is determined by a minimum goodness-of-fit measure. The method was tried on a Y-intersection. The ALV uses appropriate models derived from a map database. A second method, rather than recognizing the image road intersection, finds corridors of free space that the ALV can navigate through. A road height profile obtained from a range scanner is used to find a free space region. The height profile, representing height and normal road surface angle, is broken into solid horizontal planar segments. In a segmented image, distances to the road boundary in each of five principal directions (east, northeast, north, northwest, and west) are found for each road point. These image distances are converted to distances between the road point and the world plane boundary in each direction using the road height profile. The smallest directional distance approximates the world road boundary distance. We present results on synthetic and real road intersection images.