Automated segmentation and dimension estimation of railway bridge piers from LiDAR-SLAM point clouds

Le résumé fourni par la source

This study develops and evaluates an application-oriented integrated workflow for extracting dimensional information on railway bridge piers from three-dimensional (3D) point clouds acquired by Light Detection and Ranging (LiDAR)-based Simultaneous Localization and Mapping (SLAM). The workflow integrates bridge body extraction, pier segmentation using the SuperPoint Transformer (SPT), and cuboid-based dimension estimation. Although LiDAR-based measurement of bridge geometry has been demonstrated under controlled conditions, real-environment SLAM point clouds contain substantial noise from surrounding structures, vegetation, and registration drift, and the propagation of segmentation errors to the final dimensional estimates remains insufficiently understood for railway bridges. To address this, the workflow combines Simple Morphological Filter (SMRF)-based ground removal, height-based clustering, and principal component analysis (PCA)-driven pose correction for bridge body extraction, followed by pier segmentation using both a rule-based geometric method and the SPT. Dimension estimation is performed via Cuboid Random Sample Consensus (RANSAC) fitting, and three automation levels—fully manual, semi-automatic with the Segment Anything Model (SAM), and fully automatic—are compared to examine how errors at the extraction stage influence dimensional error. Under file/scan-level cross-validation using point clouds from two railway bridges (single-column piers and rigid-frame viaducts), the SPT achieved mean Intersection over Union (mIoU) values of 84.53% and 94.80%, respectively, which were higher than those of the rule-based method (71.83% and 78.33%) and the compared deep learning baselines. In an additional bidirectional bridge-level holdout test, the mIoU values were 73.31% and 78.82%, indicating that the file/scan-level results should not be interpreted as evidence of complete generalization to unseen bridges. The summary dimension-error value was 0.0423 m for the semi-automatic approach, while the fully automatic workflows yielded values of 0.0710–0.0913 m. These results suggest that pier extraction accuracy is an important factor influencing final dimensional accuracy within the evaluated bridge cases.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.