Recognition of Curved Objects

Abstract

In this chapter we consider the problem of recognition and localization of objects with curved surfaces. Although polyhedral objects are encountered more frequently than curved objects in an industrial environment, objects with curved surfaces are more prevalent in natural scenes. One simple way of dealing with curved objects is to approximate them by polyhedral objects—an approach commonly used in earlier vision systems. However, such an approach has certain fundamental shortcomings in that the resulting description is not rich enough to capture the intrinsic properties of curved surfaces, is not stable to changes in viewpoint, and suffers from approximation errors. In this chapter, invariant surface curvatures, viz. the mean and Gaussian and principal curvatures, are used to come up with a description of curved surfaces that is invariant to viewpoint. As in the case of polyhedral objects, multiple-object scenes with clutter and partial occlusion are considered. This precludes the use of global shape descriptors, which means that only local shape descriptors based on the invariant surface curvatures can be used. The generalized Hough transform, on account of its ease of parallelization, is chosen as the constraint propagation technique.