Detection of object wings in fused range and intensity imagery

We present methods for detecting primitives in scenes containing objects whose surfaces are approximately piecewise quadric using a sensor that yields fused range and intensity. The primitives detected in a fused image are viewpoint dependent and are called object wings, which consist of adjacent fragments of object contour and surface. Wings are small enough so that some object parts can be sensed despite significant occlusion, yet rich enough to be of use in indexing to candidate models, in constraining object pose, and in aggregation methods which combine primitives into larger structures or objects. Fitting procedures are defined for specific wing primitives. Experiments with both synthetic and real data show that classification within the wing types of planar, spherical, conical, cylindrical can be achieved with over 90% accuracy. Contour information available in fitting fused data improves fitting relative to use of surface data alone. Moreover, it is shown that only a small number of contour and surface samples are needed for extraction of primitives.