Cloud data segmentation and classification for reverse engineering using neural networks

Automatic segmentation of point data in the past has been mainly applied to single range maps. However, there is a great need for the segmentation of fully digitized objects with multiple viewpoints. This research reports on the automatic segmentation of multiple viewpoint 3D digitized data captured by a laser scanner or a CMM. This is accomplished in two steps. Firstly, the surface normal and principal curvatures are estimated at corresponding point locations. Local Darboux frame and weighted least-square surface fitting are used to calculate the normal values and curvature values of the point data. Secondly, an eight dimensional feature vector (3D coordinate, 3D normal, Gaussian and Mean curvature) is used as an input to a Self-Organized Feature Map (SOFM). A normalized feature vector and a weighted Euclidean distance are adopted in the learning process of the SOFM, which improves the speed and exactness of the segmentation. The segmentation using SOFM is robust to noise and has no limitation to surface type. The algorithm is validated by real and synthetic point data. To improve the quality of surface fitting, segmented subregions of typical surfaces are classified by using a back propagation neural network. The techniques developed play a key role in reducing the length of product development time and the quality of a final surface model.