A minimal surface perturbation method for global surface registration of unstructured point cloud data

by Aman, Ronald L.

Aman, Ronald L. A Minimal Surface Perturbation Method for Global Surface Registration of Unstructured Point Cloud Data. (Under the direction of Dr. Yuan-Shin Lee.) This thesis presents a minimal surface perturbation method for the surface registration of unstructured point cloud data. The registration problem is applicable to many disciplines. Computer vision (stereo vision), computer graphics, image processing, and reverse engineering and quality inspection are examples of areas currently implementing surface registration techniques. The surface registration problem is defined as: Given the unstructured point cloud data sets, find the rotation and translation of one data set such that the two sets are properly aligned in a single coordinate frame. Current techniques fail to find global alignment in instances where the initial starting points are not favorable. In this paper, the proposed method generates a “geometric handler” to approximate the unstructured data sets using a small number of points capturing the global shape of the data while reducing local variations or features. An iterative searching method is proposed to minimize both the surface perturbation of point cloud data and the orientation variation by using a self-organizing neural network. The transformation of the geometric handlers is found using orientation invariant methods and is applied to the original point cloud. Once the global registration is accomplished, a local registration method is invoked requiring a search in local neighborhoods only. The proposed method can be used for the global surface registration of overlapped unstructured point cloud data sets. The presented techniques can be used in reverse engineering and CAD/CAM systems for product development and surface generation from scanned data.