<p>Unmanned intelligent maritime Search and Rescue (SAR) devices are increasingly being used for monitoring and rescue operations in coastal and remote sea areas. Although existing deep learning methods are effective, their large number of parameters and high computational requirements make deployment on resource-constrained rescue devices difficult. To address this issue, we designed the GEF-YOLO model, an improved version based on YOLOv8, for detecting drowning risks at sea. This method integrates the lightweight GhostConv and replaces the C2f modules in the backbone with C2f-EMBC modules containing EMBConv, ensuring that the model remains lightweight while retaining sufficient feature extraction capability. A PConv layer is introduced before the detection head to improve computational efficiency. In addition, a Feature-Focused module is designed to integrate cross-scale features, and a stronger feature fusion capability pyramid structure, FGDPN, guided by this module, is proposed. The SEAM attention mechanism is integrated into the detection head to alleviate the impact of occlusion. Finally, the LAMP pruning method is applied to further compress the model. Experimental results show that GEF-YOLO improves accuracy by 2.8%, increases mAP by 0.7%, reduces computational cost by 44.8%, improves inference speed by 31.2%, and decreases parameter count and model size by 64.1% and 58.7%, respectively, fully demonstrating the potential of the model in drowning risk detection tasks.</p>