合成点云驱动的铁路无昨轨道点云精细化 语义分割方法研究

Point cloud semantic segmentation is the foundation for identifying and analyzing the service status of railway infrastructure. As key components densely distributed along railway lines, ballastless tracks are characterized by problems such as blurred geometric boundaries of individual components and susceptibility to interference from track surface attachments in their point cloud data. These factors cause low efficiency and high error rates in manual annotation, limiting the extensive application of deep learning in this field. To address this, this paper proposed a synthetic point cloud generation framework based on building information modeling (BIM). First, relying on the BIM model of railway ballastless tracks, synthetic point cloud data of individual components and the overall structure were obtained through phased virtual scanning. Subsequently, a hybrid index structure based on Spatial Hashing and hierarchical KD-Tree was utilized to achieve lightweight automatic semantic annotation of synthetic point clouds. On this basis, the pre-trained backbone network Swin 3D based on the Transformer model was trained by fusing measured point clouds with synthetic point clouds, applying it to the semantic segmentation task of measured point clouds collected by track 3D inspection vehicles. Experimental results show that the enhanced Swin 3D model trained by fusing synthetic and measured point clouds exhibits superior performance, with a mean intersection over union of 91. 92%. Compared with the baseline Swin 3D, the intersection of union values of key components such as spiral spikes and rail clips are increased by 6. 67% and 3. 28% respectively. In addition, this paper quantitatively evaluates the efficiency of the proposed hybrid index structure, providing a theoretical and experimental basis for the promotion and application of synthetic point clouds in railway ballastless track point cloud semantic segmentation.