A virtual testbed for maritime search and rescue: Customizing YOLO for small object detection in unreal engine

Abstract

Maritime Search and Rescue (SAR) operations face three main issues: they are time-critical, rescue boat access is often difficult, and the vastness of the ocean poses a great challenge to identifying drowning victims. Rescue operations in situations such as a shipwreck or when enthusiasts drown due to rip currents at seashores are largely unsuccessful due to an over-reliance on manual methods. The need for faster, semi-automated methods thus becomes important. This paper aims to introduce drones as agents for rapid detection and rescue in such scenarios.

This study introduces a virtual simulation testbed, created in Unreal Engine (UE) 4.2 for maritime SAR scenarios. With the integration of the Cesium plug-in, the simulation platform provides a photo-realistic representation of the Earth’s surface, making it valuable for planning search and rescue (SAR) missions. Beyond simulation, the testbed offers the potential to generate large volumes of custom synthetic data, thereby supporting future research toward safer maritime transportation.

This study introduces an object detection framework tailored for maritime SAR scenarios, employing selected models from the YOLOv11 and YOLOv12 series trained on domain-relevant classes. A systematic training approach was followed, combining dataset balancing strategies with targeted architecture modifications. These efforts led to a 38.7% improvement in the mAP₅₀ metric compared to the baseline YOLOv12s model, with the enhanced YOLOv12s achieving 94.6% mAP₅₀. Further evaluation on three edge devices demonstrated the Jetson Nano NX as the most suitable platform for deployment. Among the tested models, the enhanced YOLOv12n emerged as the optimal choice for drone-based applications, delivering 92.3% mAP₅₀ at 31.2 FPS on the Jetson Nano NX, thereby achieving an effective balance between accuracy and real-time performance.