Multi-Scale Oriented Detection with Shared Convolution for UAV-Enabled Maritime Safety Surveillance

Due to their ability to cover wide areas and adapt to variable perspectives, unmanned aerial vehicles (UAVs) equipped with high-definition cameras have become effective devices for maritime safety management. However, the changing visual angles, varying flight distances, and limited computational power pose challenges for maritime safety surveillance using UAVs. These challenges often result in inaccurate multi-angle detection, rough multi-scale vessel detection, and computational strain from large models. Therefore, we propose a lightweight multi-scale oriented detection model for UAVs. Specifically, to accommodate variable flight altitudes, we firstly proposed a cross-stage partial feature fusion module named LDFusion, which can freely adjust the size and shape of the convolutional kernel to extract and fuse features at different scales. While the LDFusion module improves feature extraction performance, it also introduces additional parameters. Therefore, we secondly designed a lightweight detection head with shared convolution module SConvs for oriented ship detection, reducing the number of parameters. Thirdly, we created 3 oriented datasets from a maritime UAV perspective, including a new inland waterway dataset, a re-annotated marine dataset, and a re-annotated complex maritime dataset. Finally, we conducted comparative experiments on the 3 datasets using advanced oriented detection methods. Experimental results demonstrate that though our method achieves a modest 3.27% improvement in detection accuracy, it reduces the number of parameters by 24.40% compared to the latest approach.