High-Precision Marine Radar Object Detection Using Tiled Training and SAHI Enhanced YOLOv11-OBB

Area of Science:

• Computer Vision
• Marine Technology
• Artificial Intelligence

Background:

• Marine radar object detection is vital for maritime situational awareness but faces challenges like sea clutter and small targets.
• Existing methods struggle with scale variability and interference in Plan Position Indicator (PPI) radar images.

Purpose of the Study:

• To develop a high-precision object detection pipeline for marine radar imagery.
• To address challenges in detecting small targets and handling scale variability in radar data.

Main Methods:

• Integration of tiled training, Sliced Aided Hyper Inference (SAHI), and an oriented bounding box (OBB) variant of YOLOv11.
• Development of a semi-automatic contour-based annotation pipeline for generating multi-format labels from radar data.
• Experiments conducted on the German Aerospace Center's (DLR) DAAN dataset.

Main Results:

• The tiled YOLOv11n-OBB model with SAHI achieved mean Average Precision (mAP@0.5) over 0.95.
• Mean center localization error was below 10 pixels, demonstrating high precision.
• The method outperformed standard baselines and other YOLOv11 variants, especially for small targets.
• Near real-time inference (4-6 FPS) achieved on edge hardware.

Conclusions:

• Oriented bounding boxes (OBBs) and scale-aware strategies significantly enhance detection precision in complex marine radar environments.
• The proposed pipeline offers practical advantages for maritime tracking and autonomous navigation systems.
• Lightweight models enable efficient, high-performance object detection on edge devices for real-time applications.