Automated multi-scale dimensional measurement of railway bridges from 3D point clouds via integration of deep learning and geometric processing

Building structural models that reflect bridge geometry is important for the analysis of railway structures such as seismic-damage simulation. This study examines a method for automatically measuring major bridge dimensions from three- dimensional point cloud data. The input consists of point clouds segmented in advance by a Superpoint Transformer-based model. We then combine these segmented point clouds with geometric processing, including principal component analysis (PCA), plane fitting, clustering, and cuboid fitting, to estimate both global dimensions and local member dimensions. The method is evaluated on 18 dimensional items for the rigid-frame viaduct and seven for the single-column pier within a common framework. Of these, 16 and six items, respectively, had corresponding design-reference and observable point-cloud extents. For these comparable items, the mean item-level absolute error ratios were 4.39% for the rigid-frame viaduct and 2.14% for the single-column pier. The workflow was evaluated on one rigid-frame viaduct and one single-column-pier structure using ten and nine repeated point-cloud acquisitions, respectively, obtained under different acquisition conditions. The results demonstrate the feasibility of the workflow for the two investigated structures and characterize repeatability and sensitivity to acquisition conditions.