Real-Time Interpretation of Typical Artificial Targets Using ARV Forward-Looking Sonar

Accurate recognition of sonar image targets is cru-cial for an Autonomous Remotely Operated Vehicle (ARV) to execute tasks in complex underwater environments. However, the ARV platform requires real-time information exchange for successful task completion. This paper develops and implements an efficient real-time sonar image processing algorithm using forward-looking sonar. By employing an improved bilinear inter-polation method, the sonar image generation rate is enhanced. The YOLOv8 model, deployed on NVIDIA's Jetson AGX Orin developer suite, facilitates real-time recognition of typical under-water artificial targets such as pipelines, cables, and suspended acoustic arrays. Experimental results demonstrate that this model achieved a precision of 95.5%, a recall of 97%, and a mean average precision of 98.2%, respectively. The pool experiment demonstrates that the system can process 8 frames of sonar images per second and upload the detection and recognition results to the ARV, meeting real-time requirements and providing a strong guarantee for the completion of underwater operations.