Abstract: Railway slope supporting structures are prone to cracks, deformation, and other defects during long-term service. Accurate semantic segmentation of point clouds of slope supporting structures is an important step for defect identification and condition assessment. However, the semantic segmentation accuracy of point clouds of slope supporting structures is limited by interference from background objects such as vegetation and slope surfaces, as well as significant variations in structural scale and complex geometric characteristics. To address these challenges, a semantic segmentation method for point clouds of point clouds of railway slope supporting structures in complex mountainous scenes is proposed. Point cloud data are acquired using a UAV-mounted LiDAR system, and ground and non-ground points are separated by Cloth Simulation Filtering （CSF）. A point cloud dataset containing retaining walls and frame beams is then constructed, and a block partition strategy is adopted to enhance sample diversity. A semantic segmentation model based on Point Transformer V3 is established and compared with representative methods, including PointNet, PointNet++, DGCNN, and Point Transformer. The results show that the proposed method achieves an overall accuracy （OA）, mean accuracy （mAcc）, and mean intersection over union （mIoU） of 94.89%, 94.58%, and 87.93% on the test set, respectively. Under complex background interference conditions, the proposed method demonstrates significantly better boundary preservation and detail recognition capabilities for supporting structures than the compared methods. The results provide a reliable data foundation and technical support for the automated extraction and quantitative analysis of railway slope supporting structures.