Date of Award
Doctor of Philosophy in Oceanography
This thesis examines ways to improve the quality of bathymetric maps generated from multibeam sonar data. In particular, it focuses on techniques for improving the real time computation of bathymetric models and related navigation solutions. Currently, there are many computational limitations that render applications such as path planning and simultaneous localization and mapping (SLAM) unfit for real time use on vehicles in the field. Much of the work presented here focuses on re-framing previously proposed mapping and navigation solutions in a way conducive to massively parallel processing on a graphics processing unit (GPU).
Terrain models produced from multibeam sonar data are typically generated by gridding methods that divide the survey area into grid cells and compute the average depth value of the points that fall in each cell. To generate a gridded terrain model that is smooth and free of gaps, the cell size needs to be sufficiently large to contains several points. A larger cell size will, however, reduce the effective resolution of the model. By stochastically modeling the terrain elevation as a two dimensional function of position, a Gaussian Process Regression (GPR) is able to compute a continuous surface that represents the data at all (x,y) positions without reducing the effective resolution while simultaneously estimating the uncertainty of the model. Despite its predictive power, GPR methods are generally relegated to post processing due to the high computational cost. The main contribution of this thesis focuses on developing, implementing, and testing a formulation of a massively parallel GPR (MP-GPR). This implementation can be computed in real time for high data rate multibeam sonars and be recursively updated as new data becomes available.
Most underwater vehicles lack the ability to precisely georeference themselves. Maps generated by these vehicles are thus limited by their navigation solution more than the precision of their perceptual sensors. However, using SLAM methods, it is possible to use perceptual data, such as multibeam sonar, to inform the navigation solution. A bathymetric implementation of a Rao-Blackwellized particle filter known as BPSLAM was previously confirmed to work using a GPR terrain model. The forking nature of the BPSLAM method is particularly compatible with the recursive nature of the massively parallel GPR algorithm for both computation and efficient data storage. A new GPU based BPSLAM (GP-BPSLAM) was developed. GP-BPLAM was able to function in real time using live multibeam sonar data and produced more self consistent maps than dead reckoning navigation alone.
Results within this thesis were obtained using data collecting with a surface vessel equipped with a suite of navigation sensors and two multibeam sonars. The presented results demonstrate the utility of the developed methods in realistic operating situations for autonomous surface and underwater vehicles.
Krasnosky, Kristopher, "MASSIVELY PARALLEL STOCHASTIC TERRAIN MODELS IN UNDERSEA MAPPING AND NAVIGATION" (2021). Open Access Dissertations. Paper 1289.