Abstract
In the field of computer vision, extraction and integration of appropriate features on surface shapes play important role of recognition processes. As such features, surface curvatures have been widely used since they reflect essential properties of 3-D objects. In this paper, we propose a method for estimating surface curvatures by means of surface fitting to range images. Furthermore, a clustering method using a Mixture Probability Algorithm is proposed for image segmentation, in which the surface is classified into regions with similar curvatures. As the approximating surface, B-spline surface with controllable knots is chosen, which can represent arbitrary 3-D surface shapes by arranging optimal knots. In order to realize the optimal knots arrangement, a knot control strategy based on recursive knot insertions, deletions and transfers is developed. The effectiveness of the proposed method has been confirmed through experiments using synthetic images and real range images.
