Surface characterization using a gaussian weighted least squares technique towards segmentation of range images

Range images provide direct depth information of object points in the scene. They provide useful 3-D information useful for object location/recognition tasks. The primary focus of this research is to develop reliable and efficient algorithms for feature extraction, surface characterization, and segmentation.

The most effective technique, in feature enhancement, is to retrieve the derivatives from a second order polynomial locally fitted via an Equally Weighted Least Squares (EWLS) technique. This approach provides smoothing but has the dis- advantage of providing poor approximations near the discontinuities. To gain higher accuracies, we develop a Gaussian Weighted Least Square (GWLS) technique that provides an additional focus element for the central pixel location. The computational efficiency was also maintained through the use of weighted sets of orthogonal polynomials. Two segmentation algorithms were developed that rely heavily on this development.

In the second segmentation algorithm, a general procedure was developed to characterize arbitrary shaped regions in range images. The proposed module minimizes the L^p norm, and its efficiency and robustness properties were evaluated against the Heavy Tailed Distribution, which is a natural extension to the normal distribution. The theoretical and experimental evidence shows the advantages of this approach over the least squares and Chebychev approximations in fitting arbitrary shaped contaminated regions.