According to the Convention on the International Regulations for Preventing Collisions at Sea (COLREGs), vessels must maintain an effective lookout by sight, hearing, and all available technical means, taking into account prevailing circumstances and environmental conditions. In this context, automated visual surveillance systems represent an important enhancement of conventional navigational tools, enabling early detection and interpretation of surrounding vessel behaviour. This study proposes a video surveillance system based on the lightweight YOLOv8n deep learning model for the detection and classification of eight vessel aspects. The model was trained on an initial dataset of 925 images, which was further expanded to 1,742 annotated images. The dataset was designed to reflect real maritime operating conditions, including different times of day, weather scenarios, vessel categories, and geographical regions. To improve robustness, data augmentation techniques such as colour space transformations, geometric modifications, and classification-specific augmentations were applied. Class imbalance was mitigated through the use of class weighting. The paper also describes the system architecture and camera configuration, providing effective surveillance coverage up to 6 nautical miles. The proposed approach enables not only vessel aspect recognition but also the estimation of relative and true bearings, thereby contributing to improved situational awareness and collision avoidance in maritime navigation.
Pashenko et al. (Thu,) studied this question.