Research on Two-Stage Railway Intrusion Detection Using Enhanced U-Net and YOLOv11

In the domain of railway foreign object intrusion detection, full-frame approaches with original algorithms suffer from high false-alarm rates and poor sensitivity to small targets in complex backgrounds. To overcome limitations, this paper presents an innovative two-stage detection framework based on an enhanced U-Net and YOLOv11. In the first stage, a segmentation network extracts the track area as the region of interest by integrating a Dense Connected Dilated block and a Non-Local block into the U-Net architecture, thereby enriching multi-scale feature extraction and embedding global attention. In the second stage, a dual-branch Spatial-Channel Gating block is incorporated into the YOLOv11 backbone to suppress background noise and refine feature fusion. On RailGoerl24 datasets, the augmented segmentation model raises mIoU from 86.6% to 92.0%, achieving precise track delineation, while the modified detector attains 73.0% mAP and sustains real-time inference at 58 FPS. Systematic experiments validate that the proposed system significantly enhances small-object detection accuracy and maintains high processing speed, offering a robust and deployable solution for real-world railway safety applications.