AN AUTONOMOUS OBSTACLE AVOIDANCE METHODOLOGY FOR UNCREWED SURFACE VEHICLES FACILITATED BY DEEP-LEARNING BASED OBJECT DETECTION.

Uncrewed surface vehicles (USVs) have gained significant attention in the past decade due to their potential commercial applications and a vast global market. An essential capability of USVs is autonomous obstacle avoidance in oceanic environments to prevent collisions. This study focuses on investigating and benchmarking deep learning-based real-time object detection algorithms for collision avoidance in USVs. The candidate algorithms were trained and evaluated using a maritime dataset provided by Marine Thinking Inc. Among the algorithms tested, You Only Look Once (YOLO) v5m demonstrated outstanding performance (mAP50 > 0.84 and mAP50:95 > 0.69) and high frame per second (FPS > 60) in realtime scenarios. Subsequently, the pre-trained YOLO v5m algorithm was converted into a TensorRT engine and deployed within a DeepStream pipeline on an onboard Nvidia Jetson for real-time collision avoidance tests. The results indicate that the YOLO v5m algorithm successfully detected obstacles and provided feedback to the ArduPilot autonomous suite. The ArduPilot, in turn, adjusted the control signal of the USV motor based on the Bendy Ruler obstacle avoidance algorithm, resulting in the successful avoidance of detected obstacles, and preventing collisions. This study presents an efficient, production-ready, and cost-effective autonomous collision avoidance methodology for USVs. [ABSTRACT FROM AUTHOR]