Abstract
Introduction The detection of occluded marine vessels is critical for the safe navigation and operation of unmanned surface vehicles (USVs). While image-based detection methods have achieved substantial accuracy, their high computational and memory requirements prohibit deployment on resource-constrained embedded platforms. To address this, we propose eAodeMar (efficient AodeMar), a lightweight version built upon our prior AodeMar model, specifically designed for efficient occluded marine vessel detection. Methods The efficiency of eAodeMar is achieved by integrating Ghost convolution modules into both the backbone and the feature fusion network, significantly reducing model parameters and computational load while maintaining accuracy. To ensure practical applicability, the optimized model is deployed on an embedded GPU platform (Jetson Xavier NX), incorporating dedicated structural refinement and inference acceleration techniques. Results Extensive experiments on the public MVDD13 dataset demonstrate that eAode- Mar reduces parameter count and computational load by 7.00% and 0.89%, respectively, with only a marginal accuracy drop of 0.42%, while achieving a remarkable 42.12% improvement in inference speed. When deployed on the Jetson Xavier NX device, it attains a real-time detection rate of 28.57 FPS on the SMD video stream. Discussion These comprehensive results validate that eAodeMar effectively balances high precision with high efficiency in occlusion-prone maritime environments. The model demonstrates strong potential for real-world ocean engineering applications, offering a practical solution for real-time detection on embedded systems.