| Abstract | For the safe and reliable navigation of USVs (Unmanned Surface Vehicles), it is essential that they autonomously recognize maritime obstacles and accurately detect their surroundings. To achieve this, USVs are typically equipped with multiple sensors, such as RADAR (RAdio Detection And Ranging) and a camera, each possessing distinct advantages and limitations. RADAR provides long-range detection, relative speed measurement, and stable performance under various lighting and weather conditions. However, its low resolution reduces accuracy in object classification and lateral distance estimation. In contrast, the camera offers high-resolution visual data for detailed recognition but is sensitive to lighting conditions and generates large amounts of information that require significant processing. Because of differences in sensing range, frequency, and error characteristics, achieving consistent, accurate obstacle tracking with a single sensor is challenging. Therefore, it is necessary to integrate multiple sensors, combining RADAR's robustness across various weather conditions with a camera's high-resolution capabilities. To address these challenges, we propose an improved method for detecting and tracking maritime obstacles using multiple-sensor fusion. The proposed framework improves overall reliability in two steps. First, the tracking performance of each sensor—RADAR and the camera—is enhanced to ensure that both sensors can independently deliver accurate, stable tracking results. Then, the two sensors are combined through sensor fusion to complement each other, enabling the system to achieve higher accuracy and stability in tracking maritime obstacles. The framework consists of three main stages: (1) detection and localization, (2) tracking, and (3) sensor fusion. In the detection and localization stage, the camera detects obstacles using the YOLO (You Only Look Once), a representative deep learning-based object detection model, and a depth-based localization method is introduced to mitigate the instability of the conventional horizon-based localization method. For RADAR, obstacle detection and localization are performed using the CA-CFAR (Cell-Averaging Constant False Alarm Rate) algorithm, a representative RADAR signal-processing method that ensures robust detection even in cluttered maritime environments. In the tracking stage, a hybrid tracking method is developed by combining the EKF-based (Extended Kalman Filter) models with a learning-based model. This hybrid approach leverages the interpretability and stability of EKF-based tracking and the adaptability of learning-based tracking, providing robust performance across varying maritime environments. Finally, in the sensor fusion stage, the tracking results from RADAR and the camera are integrated using a sensor-level fusion method based on the estimated error covariance, yielding accurate and stable tracking performance. We validated our method through field experiments using RADAR and camera sensor measurements. The results demonstrate that the proposed method significantly improves detection accuracy and tracking stability compared to single-sensor approaches. These results confirm that the proposed framework effectively enhances situational awareness and navigation safety for USVs operating in the complex maritime environments. |
|---|---|
| Publication Date | 2026-06-09 |
Yun-Sik Kim, Myung-Il Roh, Ha-Yun Kim, In-Chang Yeo, Nam-Sun Son, "An Improved Method for Detection and Tracking of Maritime Obstacles Using Multiple-Sensor Fusion", Proceedings of OMAE (International Conference on Ocean, Offshore and Arctic Engineering) 2026, Tokyo, Japan, 2026.06.07-12