3D Road Scene Interpretation for Autonomous Vehicle Driving

In this paper, the problem of 3D road scene interpretation for autonomous vehicle driving is addressed. In particular, the problems of road detection and obstacle avoidance in outdoor environments are investigated. A set of descriptive primitives (straight and circular line segments) is selected to describe 3D objects which commonly occur in road scenes, e.g., people, cars, trucks, houses, etc. First, these primitives are extracted directly from the input image of the scene, and then are grouped according to specific geometric relationships (symmetry, convergence, parallelism, closeness, etc.). Relational geometrical knowledge of the elements of a group can be used to index an object in a pure bottom-up way, so decreasing the recognition complexity by reducing the amount of data to be matched with an object model database. Results on a road image containing obstacles, which show the efficiency, accuracy and time performances of the proposed method are reported