Abstract: Realizing high-precision semantic segmentation of railroad scenes can optimize asset management, improve planning and design, and advance railroad intelligence. In response to the problems of high data acquisition cost, low segmentation accuracy, and poor generalization ability in current railway scene semantic segmentation, this paper proposes a railway scene semantic segmentation method based on an improved dynamic graph convolutional neural network (SAM-DGCNN). The method first uses UAV to acquire multi-view images of railroad scenes; then generates 3D point clouds by SFM-PMVS algorithm; and finally finishes semantic segmentation of point cloud data by improved graphical convolutional neural network. The results of the study show that SAM-DGCNN network based on image point cloud has higher accuracy for semantic segmentation of railroad scenes, and the segmentation accuracy is 4% and 6% higher than that of DGCNN network and PointNet++ network, respectively.