Abstract: A railway scene semantic segmentation method based on improved dynamic graph convolutional neural network is proposed to address the issues of high data acquisition cost, low segmentation accuracy, and poor generalization ability in current railway scene semantic segmentation. Firstly, high-resolution drones are used to capture multi view images of railway scenes, and 3D point clouds of railway scenes are generated through structural motion recovery and patch based multi view stereo vision algorithms. Then, a spatial attention module is introduced into the dynamic graph convolutional neural network to enhance the segmentation accuracy and generalization of the network structure. Finally, an improved graph convolutional neural network is used to achieve high-precision semantic segmentation of the pre-processed railway scene point cloud. The railway scenes used in the experimental stage include bridge sections, roadbed sections, and connecting lines, totaling 11 areas. Compared with dynamic graph convolutional neural networks and PointNet++, the research results show that the improved dynamic graph convolutional neural network based on image point cloud training has higher accuracy in semantic segmentation of railway scenes, with an average intersection over union ratio as the evaluation indicator. Compared with dynamic graph convolutional neural neural networks and PointNet++, the segmentation accuracy is improved by 3.3 percentage points and 6.0 percentage points, respectively, and it has better generalization ability.