SN-YOLO: A Heavy-Haul Railway Perimeter Intrusion Detection Method Based on YOLOv11

As a national freight backbone network, heavy-haul railways play a vital economic backbone role. However, the absence of protective measures in certain sections combined with the inability to brake rapidly renders these railways highly susceptible to severe pedestrian intrusion accidents. This paper proposes an improved YOLOv11-based small target detection framework for UAV-based heavy-haul railway inspection. The framework incorporates a Space-to-Depth Convolution feature extraction module, which replaces traditional strided convolutions and pooling layers. This module preserves fine-grained features of small targets while reducing feature map resolution, effectively preventing information loss during downsampling. Additionally, the Normalized Wasserstein Distance loss function is adopted, modeling bounding boxes as Gaussian distributions. This approach utilizes the normalized Wasserstein distance to measure similarity between predicted and ground-truth bounding boxes, resolving anchor matching failures caused by IoU’s sensitivity to minor positional variations in small targets and significantly enhancing detection stability. The experimental results show that validation on the self-constructed Beijing-Ma’anshan heavy-haul railway dataset achieved 73% mAP and 76.4% precision, demonstrates that the SN-YOLO method accurately identifies minuscule pedestrian targets within complex heavy-haul railway scenarios, overcoming target blurring and feature dominance by larger objects. This provides a robust technical solution for real-time perimeter security monitoring in railway systems.