Date of Award
Master of Science in Ocean Engineering
Massively Parallel Gaussian Process Regression (MP-GPR) is a parallelized programmatic implementation of a machine learning algorithm used to process multi-beam sonar data and generate 2.5D terrain maps in close to real-time. This thesis furthered the capabilities of MP-GPR in two ways. The first introduced data reduction techniques to reduce the quantity of data inputted to the MP-GPR algorithm for computational efficiency. Various approaches were analyzed to determine an optimal data minimization strategy. The second informed the variances of regression estimates by considering beam angle and range uncertainty of the multi-beam sonar data being analyzed. In tandem, the presented methodologies decreased the GPR algorithm run-time and provided an improved uncertainty model. Trials were run using a Graphical Processing Unit (GPU) to process data from a Robot Operating System (ROS) data bag with C++. The data was obtained from a WASSP multi-beam sonar survey of the Wiggles Bank on St. Mary's River in Georgia, USA. The results showed that there were no significant gains in accuracy by using more complex downsampling algorithms and instead using simple method (systematic, average, hybrid) was adequate.
Parisi, Phillip M., "FURTHERING USABILITY AND EFFICIENCY OF MASSIVELY PARALLEL GAUSSIAN PROCESS REGRESSION AS APPLIED TO MULTI-BEAM SONAR DATA" (2023). Open Access Master's Theses. Paper 2322.