Second-order implicit polynomials for segmentation of range images

Abstract In this paper, we present an algorithm for segmentation and feature extraction for noisy range images. We use the edge- and region-based approach for obtaining reliable segmentation maps. For edge detection, regions having much higher values of planar fit error are assumed to contain edges. These edge regions are explored for detection of crease edges, which can be located at the maxima of the curvature. However, due to noise, the maximas are generated at a number of points other than the actual edges. These extraneous curvature maximas have a substantially smaller value of curvature compared with the curvature of the actual edges. The true edges are detected using an approach based on local thresholding within edge regions. The detected edges segment the range image into smooth patches. On each of the smooth patches, a planar/quadric fit is obtained using a novel non-iterative procedure. To obtain reliable fits, we had to restrict the permissible class of quadrics to surfaces of revolution. The quadric fit obtained gives a very reliable estimate of the position and orientation of the axis for curved surfaces. The approach has been demonstrated on real range data and has been found to give good results.