Automatic segmentation of digitized data for reverse engineering applications

Reverse engineering is the process of developing a Computer Aided Design (CAD) model and a manufacturing database for an existing part. This process is used in CAD modeling of part prototypes, in designing molds, and in automated inspection of parts with complex surfaces. The work reported in this paper is on the automatic segmentation of 3-Dimensional (3-D) digitized data captured by a laser scanner or a Coordinate Measuring Machine (CMM) for reverse engineering applications. Automatic surface segmentation of digitized data is achieved using a combination of region and edge based approaches. It is assumed that the part surface contains planar as well as curved surfaces that are embedded in a base surface. The part surface should be visible to a single scanning probe (2½D object). Neural network algorithms are developed for surface segmentation and edge detection. A back propagation network is used to segment part surfaces into surface primitives which are homogenous in their intrinsic differential geometric properties. The method is based on the computation of Gaussian and mean curvatures of the surface. They are obtained by locally approximating the object surface using quadratic polynomials. The Gaussian and mean curvatures are used as input to the neural network which outputs an initial region-based segmentation in the form of a curvature sign map. An edge based segmentation is also performed using the partial derivatives of depth values. Here, the output of the Laplacian operator and the unit surface normal are computed and used as input to a Self-Organized Mapping (SOM) network. This network is used to find the edge points on the digitized data. The combination of the region based and the edge based approaches, segment the data into primitive surface regions. The uniqueness of our approach is in automatic calculation of the threshold level for segmentation, and on the adaptability of the method to various noise levels in the digitized data. The developed algorithms and sample results are described in the paper.