Towards real-time sea-floor surface reconstruction and classification using 3-D side-scan sonar

This thesis presents a computer algorithm to solve two major hurdles for generating real-time automated sea-floor maps with composition classification using 3-D side-scan sonar data. The algorithm consists of two distinct parts: sea-floor profiling and sea-flooring classification with computation acceleration from a graphics processing unit (GPU). The sea-floor profiling algorithm is an automated method that identifies bathymetry data corresponding to the sea-floor while ignoring bathymetry corresponding to water column objects and multi-path returns. The algorithm improves upon a fuzzy curve tracing method to handle discontinuities in the point-cloud data along the sea-floor and to discriminate between the sea-floor and other data. With an average error of 2.6% and a computation time of 7.40ms, the sea-floor profiling algorithm is extremely accurate and efficient. Classification of the sea-floor regions consists of applying image texture methods and machine learning classifiers to side-scan sonar images. In this thesis, a feature space for each side-scan sonar image pixel is created using image texture analysis algorithms, and classified with an artificial neural network. The accuracy and performance of the algorithm is tested with side-scan sonar images from the Underwater Research Lab's Pam Rocks sonar survey. Real-time classification was achieved by the use of GPU computing. Porting the algorithm onto the GPU using OpenCL reduced the per-ping computation time to an average of 100ms, with an average error of 3.4%, making it a viable real-time solution in a sonar system.