SLAM for Structured Environments Using Mechanically Scanned Imaging Sonar

Abstract

As the modern utilization of the maritime environment only grows, uncrewed systems present the future of safety, efficiency, and capability. For submerged operations, Autonomous Underwater Vehicles (AUVs) enable scientists, industry, and militaries to access remote, inhospitable locations and execute a variety of tasks beyond the capabilities of human occupied or operated systems. Much of this autonomy relies on the vehicle having a detailed understanding of its own position. Inertial Navigation Systems provide an estimate of the distance traveled by combining numerous sensors, but are subject to unbounded error accumulation over long distances. Traditional methods of correcting for this error found in terrestrial robotics are largely unavailable in the undersea domain due to the absorption and scattering effects of electromagnetic signals in water. Acoustic communications and imaging such as Sound Navigation and Ranging (SONAR) is the most reliable and trusted method for AUVs. This thesis presents a novel method for performing Simultaneous Localization and Mapping (SLAM) through acoustic means utilizing a Mechanically Scanned Imaging Sonar (MSIS). MSIS utilize a single beam sonar mechanically rotated around the vehicle to scan a full 360◦ area. Compared with other sonar systems of similar capabilities, they require less size, weight, and power, and are available at a lower price point. The primary contribution of this thesis is a SLAM processing pipeline from MSIS to global position estimate. The pipeline extracts information from the MSIS data regarding the vehicle’s relative location compared to observed landmarks and then probabilistically matches the observed data to a best estimate vehicle position. The system is compatible with either an a priori map or a constantly updated SLAM global map. Individual beams from the MSIS are fused together into a submap. Contrast-based image processing identifies features of interest in the submap and appropriate features are then classified as observed landmarks. A probabilistic coarse-to-fine voting scheme identifies the most likely pose of the vehicle using the global map. When performing SLAM without an a priori map, observed landmarks are then evaluated and either added to the global map or used to update the position of known landmarks. While prior works have established MSIS SLAM by focusing on a single return per sonar beam, this thesis utilizes submaps to extract numerous features from a series of consecutive beams, allowing for more detailed and comprehensive feature mapping. Experimental validation was performed using an ISS360 sonar mounted on a REMUS-100 AUV with the processing pipeline running via Robot Operating System on the vehicle backseat computer. The vehicle was assisted by divers traversing underneath the WHOI Iselin pier and performed both localization and SLAM using the submerged pier pilings. The system performed real-time localization, successfully bounding previously unbounded localization drift to an average of 3.4m, resulting in over a 90% reduction in absolute error after approximately one hour of submerged operations. The SLAM results mirrored the a priori accuracy demonstrating similar error bounds validating the system performance.