Semantic Segmentation of Railway Infrastructure based on Virtual Model Synthetic Data

The application of digitalization in railway infrastructure is steadily increasing, and it accelerates the process of automating the asset management and inspection process of railway infrastructure. In this study, a point cloud segmentation method for railway infrastructure is proposed based on virtual model synthetic data and improved dynamic graph convolutional network (DGCNN). First, virtual data of railway infrastructure is created and inserted into real data for blending and augmentation, which is used to train point cloud segmentation neural networks. Acceptable segmentation results are achieved by using improved dynamic graph convolutional neural networks to train the augmented data and using the real data for point cloud segmentation. It is shown that the dataset augmented with virtual data achieves a 3.38% improvement in the accuracy of the final point cloud segmentation over that without augmentation, and the segmentation accuracy using the improved DGCNN network improves by 2.97% over that of the DGCNN network without improvement. This work could effectively improve the effect of 3D point cloud segmentation model to achieve accurate and efficient digitization of railway infrastructure.