DeUNet-YOLO: An end-to-end framework for foggy rail obstacle detection

Reliable obstacle detection is crucial for ensuring rail transit safety, particularly under dense fog conditions, where degraded visibility significantly increases the risk of missed detections and false alarms. In this paper, we propose DeUNet-YOLO, an end-to-end fog-aware feature compensation framework for robust rail obstacle detection. Different from conventional approaches that separately perform image enhancement and object detection, the proposed framework jointly optimizes enhancement and detection through a learnable deformable UNet module. Specifically, deformable convolutions embedded in the UNet-based enhancement front-end adaptively adjust feature sampling locations to compensate for fog-induced appearance degradation and spatial distortions, while the enhanced representations are directly utilized by the YOLO detector for obstacle recognition. Extensive experiments are conducted on both synthetic foggy datasets and a real-world foggy benchmark, demonstrating that DeUNet-YOLO achieves mAP@50 values of 80.9% and 70.2%, respectively. Moreover, the proposed method maintains real-time inference capability, achieving 70.4 FPS. These results validate the effectiveness and robustness of DeUNet-YOLO for reliable obstacle detection in challenging foggy rail transit environments. Future work will focus on reducing computational complexity and improving deployment efficiency on resource-constrained railway platforms.