DC-YOLO: A Few-Shot Transformer-Enhanced YOLO for Foreign Object Detection in Suspended Monorail Systems

Foreign object intrusion poses a critical threat to suspended monorail systems, particularly for airborne and floating objects. Traditional object detection methods excessively rely on large, annotated datasets, which are scarce for such rare intrusion events, limiting their practical deployment. To address this data scarcity and generalization challenge, we propose DC-YOLO, an efficient few-shot object detection (FSOD) method for monorails based on the YOLOv11 architecture. Our method introduces two key innovations: the C3TR module and D-Loss. The C3TR module significantly improves the model's ability to capture global contextual information and long-range dependencies crucial for detecting dispersed or subtle floating objects from limited samples. D-Loss enhances bounding box regression precision and convergence for small, airborne objects. Experimental results on our SMRail Dataset, comprising 200 images with four novel object categories, demonstrate that DC-YOLO achieves a mAP @0.5 of 0.995 and a mAP @0.5:0.95 of 0.857, notably outper-forming state-of-the-art baselines. Our method provides a lightweight, real-time capable, and robust solution for enhancing monorail safety from few-shot samples.