CONTEXT-CONSTRAINED MATCHING OF SCENE INTERPRETATION HIERARCHICAL CAD-BASED MODELS FOR OUTDOOR

We propose a method for outdoor scene interpretation which recognizes many kinds of objects in the scene, and determines their geometric structures using context information of objects in the scene and flexible matching of hierarchical CAD-based models. Unlike conventional CAD-based vision systems, our system can deal with outdoor scenes with artificial objects which may have various shapes and sizes, and natural objects for which there are no CAD-based models. The features of our system are that we effectively use the intermediate results of the scene interpretation as context information to prune the candidate object classes, their positions and orientations, and that we use hierarchical CAD models so that higher models may be able to represent one object class such as house or automobile with several variations by allowing cha.nges of dimensions. First of all, the input scene is segmented into a ground plane and objects on it using the multisensory information such as video and range images, and planar surface patches are extracted from the segmented regions. Next, flexible matching between the scene descriptions and the hierarchical CAD models are performed. Although there are many DOFs, we can constrain the range of the object class and its attitude using context information from the scene interpreta-, tion. Then, we can identify each object in the scene and infer its geometric structure. The preliminary results applied to complex road scenes are shown.