Abstract
Conventional methods for localizing nearby other ships using multiple sensors have faced limitations in practical applicability in real maritime environments due to high costs and environmental constraints. To address this, the present study proposed a method that accurately localizes other ships using only a camera. First, a virtual maritime environment was constructed to precisely replicate a real maritime environment, enabling efficient reproduction of diverse maritime scenarios and the generation of a large-scale dataset through automatic data labeling. Based on this dataset, we trained MonoLSS (Mono-Learnable Sample Selection), an image-based 3D object detection model, and used it to localize other ships. To evaluate the performance of the proposed method, we compared it with existing horizon-based and depth-based methods for localization across five representative scenarios: head-on, following, crossing with a single ship, crossing with multiple ships, and heavy storm. The results show that the proposed method achieved MAEs of 2.39 m, 2.20 m, 2.09 m, and 2.31/2.66 m (for each ship in the crossing scenario), and 3.28 m across the five scenarios. These results correspond to an error reduction of up to 92.03 % compared with conventional methods. The results demonstrate that the proposed method enables robust, real-time localization of other ships using a single camera, regardless of environmental conditions. Therefore, it is expected to make a significant contribution to the realization of safe autonomous navigation in real maritime environments where the use of additional sensors is limited.
