Sonar-equipped unmanned surface vehicle for search and rescue operations

Abstract

Drowning remains a major public safety challenge and accounts for hundreds of thousands of fatalities globally, with survival rates heavily dependent on rapid response. Current underwater search and rescue (SAR) methods, including rescue divers, remotely operated vehicles, and side-scan sonar systems, are often resource intensive, slow to deploy, or otherwise operationally limited. This thesis investigates whether a lightweight, rapidly deployable and sonar-equipped unmanned surface vehicle (USV) can provide a faster and more efficient solution for underwater SAR operations. Building upon the ongoing Seadragon initiative within the Division of Vehicle Engineering and Autonomous Systems at Chalmers University of Technology, a redesigned USV platform (Seadragon 2.0) was developed with improved structural durability and sensing capability. The prototype integrates a Teledyne BlueViewM900 Mk2 multibeam imaging sonar and an NVIDIA Jetson Orin Nano edge-computer within a compact catamaran-style hull architecture. To enable automatic underwater target detection, a custom sonar dataset was collected using recordings of submerged humans and simulated underwater debris. A two-stage transfer learning pipeline based on the YOLO26n architecture was trained both on the custom dataset and a larger public sonar dataset. While maintaining a real-time inference performance of 39.1 FPS on the embedded platform, the resulting model achieved a precision of 0.868, recall of 0.812, and mAP of 0.859. In addition, the redesigned Seadragon prototype achieved an estimated theoretical search coverage rate of 244.3 m2/s, exceeding that of conventional rescue divers and comparable side-scan sonar vessel operations. Although full-scale field validation was limited by project time constraints, the results demonstrate the feasibility of the lightweight USV concept for enhancing underwater SAR operations and, critically, reducing time to rescue.