An improved method of detecting the horizon for tracking maritime obstacles

For safe navigation of ships, it is essential to accurately detect and continuously track surrounding obstacles. The horizon, serving as the boundary between the sky and the sea, plays a crucial role in enabling ships to maintain precise routes and effectively assess the positions of obstacles. Reliable horizon detection is, therefore, highly significant. Conventional horizon detection methods, such as the Hough transform and edge-based approaches, have shown good performance in relatively simple environments. However, their accuracy significantly decreases in realistic maritime environments, which involve complex factors such as terrain, obstacles, and reflections. To address these limitations, this study proposes a novel horizon detection method that fine-tunes the SAM (Segment Anything Model), a deep learning model specifically designed for maritime images. An efficient adapter-based fine-tuning technique was implemented on the SAM' mask decoder, enabling the model to effectively learn the distinct visual and structural characteristics of maritime environments. Experimental evaluations demonstrated that the method combining the SAM fine-tuning with the vertical edge response approach achieved superior performance, significantly reducing height error and slope error by an average of 75.58% and 70.17%, respectively, even in highly complex environments. These findings highlight the superior accuracy and robustness of the proposed method, indicating its substantial potential for practical applications in autonomous navigation systems and enhanced maritime safety.