Autonomous Rail Rapid Transit Pavement Damage Segmentation Algorithm Based on YOLOv8s-D

The integrity of autonomous rail rapid transit (ART) is crucial for efficient transportation system operation. However, accurately identifying damage to ART pavements in complex backgrounds poses a challenge for existing semantic segmentation algorithms. To address this, we introduce you only look once (YOLOv8s-D), a model based on YOLOv8s with enhanced features. Modifications include an upgraded backbone network, incorporation of the bilevel routing attention (BRA) mechanism, and optimized Neck layer using dynamic snake convolution (DSConv). Specifically, the dilated dense convolutional network v4 (DDCNv4) module enhances adaptability to damage features of varying scales, and BRA boosts recognition accuracy under complex illumination and occlusion. DSConv reduces model parameters and accelerates inference speed. The model is trained and validated on a 16,000-image data set from automated drone imaging. YOLOv8s-D outperformed YOLOv8s, UNet, DeepLabV3, and lightweight reduced atrous spatial pyramid pooling, achieving a 92.7% mean intersection over union (MIoU), a 5.9% increase compared with YOLOv8s. Ablation experiments confirmed the efficacy of each enhanced module. Evaluation on other data sets revealed YOLOv8s-D’s robust generalization. A graphical user interface (GUI) facilitates inspector use. In conclusion, YOLOv8s-D enhances segmentation accuracy while minimizing model parameters, improving deployability on mobile devices and providing an efficient solution for detecting pavement damage in ART systems.