Addressing the challenge of low-light conditions and the coexistence of multi-scale targets in complex heavy-haul railway environments, this paper proposes an enhanced YOLOv10-based algorithm for foreign object detection. Our design incorporates a three-fold optimization strategy to boost model performance. First, a Preprocessing Enhancement Network is integrated into the input layer, leveraging feature decomposition and adaptive correction to enhance the visibility of object contours and textures under low illumination. Second, we introduce a Reparameterized Convolution Optimization module to minimize computation via multi-branch training and single-branch inference. Third, an improved Haar Wavelet Downsampling module utilizes frequency domain decomposition to effectively preserve detail information for small objects. Trained and validated on the OSDaR23 railway dataset, the improved model achieves a mAP of 87.2%, representing a 4.9% increase over the YOLOv10n baseline. The model maintains real-time monitoring capability with 58 FPS and 9.3 GFLOPs. Furthermore, its mAP outperforms YOLOv8n, Fast R-CNN, and YOLOv5n by 19.5%, 15.8%, and 28.2%, respectively. This research provides reliable technical support for intelligent security systems in heavy-haul railways.