Turnout point cloud segmentation method based on multi-scale fusion

SONG Yixiao; ZHAO Xinxin; WANG Shengchun; YAN Zhicheng; LI Qingyong

doi:10.11860/j.issn.1673-0291.20240014

Abstract：To address the limitations of current turnout track inspection methods,such as heavy reli-ance on manual labor,low detection efficiency,and the lack of depth information in 2D visual ap-proaches,this paper proposes a turnout point cloud segmentation method based on a multi-scale fusion strategy,named Point-Bidirectional Encoder Representations from Transformers-Turnout(Point-BERT-T).First,in the local point cloud encoding stage,spherical groupings with varying radii are employed to extract and fuse features from points within each sphere,generating a spatially hierarchi-cal hybrid feature representation.This multi-scale fused feature captures information at different spatial levels of the turnout,improving the efficiency and accuracy of railway infrastructure recognition and segmentation.It significantly enhances the capability of 3D point cloud recognition for turnouts and supports downstream tasks such as defect and deformation detection.Next,a random rotation-translation and non-uniform slicing strategy is introduced during data preprocessing to simulate real-world scanning variability,thereby improving the model's robustness under diverse data acquisition conditions.Finally,to validate the effectiveness of the proposed method,comparative experiments are conducted against existing approaches.The research results demonstrate that,compared to Point-BERT,the Point-BERT-T method improves the overall turnout point cloud segmentation perfor-mance by 1.9％.Furthermore,for the more challenging frog and wing rail components,the Intersec-tion over Union(IoU)increases by 4.7％and 5.6％,respectively,demonstrating the method's accu-racy and robustness in semantic segmentation of 3D turnout point cloud data.