Research on Rotated Target Detection Method for Maritime Vessels in UAV Aerial Images Based on Deep Learning

Abstract

Aerial imagery captured by unmanned aerial vehicles (UAVs) over maritime scenes is characterized by a wide field of view, homogeneous backgrounds, and the presence of densely packed vessels with diverse orientations. Traditional horizontal bounding box detection methods applied in such scenarios often introduce excessive background noise and fail to accurately represent vessel orientations, resulting in inadequate differentiation of densely moored vessels. To address this issue, this paper proposes an improved rotated object detection method. Building upon the YOLOv8-OBB model with inherent rotation detection capability, the proposed method incorporates a Coordinate Attention (CA) module to enhance feature extraction for elongated vessel targets, optimizes the Feature Pyramid Network (FPN) structure with cross-level skip connections to strengthen multi-scale feature fusion, and employs a Skew Intersection over Union (SkewIoU) loss function to improve the regression accuracy of rotated bounding boxes. For comprehensive performance evaluation, experiments are conducted on the public maritime dataset HRSC2016 and a self-constructed nearshore vessel subset. Results indicate that the improved model achieves a mean Average Precision (mAP@0.5) of 90.5% on the HRSC2016 test set, representing a 1.3 percentage point improvement over the baseline model (89.2%). Ablation studies demonstrate the consistent contributions of each enhancement module, and the model exhibits significantly improved robustness in detecting dense and small-scale vessels. This work provides a more accurate solution for maritime surveillance, search and rescue, and related applications.