A Fiducial Marker System for ID Recognition in Forward-Looking Sonar Images

We present a fiducial marker system tailored for underwater acoustic imaging, enabling accurate detection and recognition of multiple marker IDs in real-world Forward- Looking Sonar (FLS) images. The marker is physically de- signed with layered concrete–metal structure to generate strong and distinctive sonar reflections. Our marker detection and recognition pipeline is trained entirely on simulation data, yet it achieves accurate performance on real-world sonar images. By leveraging a custom FLS simulator we generate annotated training samples that closely mimic real sonar characteristics. A YOLO-based detector, trained with these simulated images, localizes markers and regresses corner keypoints. For marker identity recognition, detected regions are rectified and decoded using a grid-based binary recognition scheme. Experiments show that the model achieves a 86.6% true positive detection rate and 100% ID recognition accuracy in the fully visible patch subset of real sonar images, despite being trained solely on synthetic data. This sim-to-real framework offers a scalable solution for underwater localization and inspection in autonomous robotic systems.