Date of Award

12-2007

Document Type

Thesis

Degree Name

Master of Science (MS)

Legacy Department

Mechanical Engineering

Advisor

Kurfess, Thomas R

Abstract

Metrology systems take coordinate information directly from the surface of a manufactured part and generate millions of (X, Y, Z) data points. The inspection process often involves fitting analytic primitives such as sphere, cone, torus, cylinder and plane to these points which represent an object with the corresponding shape. Typically, a least squares fit of the parameters of the shape to the point set is performed. The least squares fit attempts to minimize the sum of the squares of the distances between the points and the primitive. The objective function however, cannot be solved in the closed form and numerical minimization techniques are required to obtain the solution. These techniques as applied to primitive fitting entail iteratively solving large systems of linear equations generally involving large floating point numbers until the solution has converged. The current problem in-process metrology faces is the large computational times for the analysis of these millions of streaming data points. This research addresses the bottleneck using the Graphical Processing Unit (GPU), primarily developed by the computer gaming industry, to optimize operations. The explosive growth in the programming capabilities and raw processing power of Graphical Processing Units has opened up new avenues for their use in non-graphic applications. The combination of large stream of data and the need for 3D vector operations make the primitive shape fit algorithms excellent candidates for processing via a GPU. The work presented in this research investigates the use of the parallel processing capabilities of the GPU in expediting specific computations involved in the fitting procedure. The least squares fit algorithms for the circle, sphere, cylinder, plane, cone and torus have been implemented on the GPU using NVIDIA's Compute Unified Device Architecture (CUDA). The implementations are benchmarked against those on a CPU which are carried out using C++. The Gauss Newton minimization algorithm is used to obtain the best fit parameters for each of the aforementioned primitives. The computation times for the two implementations are compared. It is demonstrated that the GPU is about 3-4 times faster than the CPU for a relatively simple geometry such as the circle while the factor scales to about 14 for a torus which is more complex.

Share

COinS