3D point cloud classification of bridges using superpoint transformer

Abstract

Automatic component classification was performed on bridge point clouds acquired using LiDAR (Light Detection and Ranging). The effects of voxel size in the preprocessing stage and input features on classification performance were systematically analyzed using the Superpoint Transformer (SPT).A comparison of voxel sizes (0.20, 0.10, 0.07, and 0.05 m) showed that the highest mean Intersection over Union (Mean IoU) was achieved at a voxel size of 0.07 m (83.15%). When the point cloud was downsampled with a voxel size of 0.20 m, the IoU values for the deck slab and parapet decreased markedly. Furthermore, under the same experimental conditions as those in previous studies, the Mean IoU improved from 74.43% to 83.15%, demonstrating an overall enhancement in classification accuracy.In addition, geometric features computed using geometric features (e.g., Linearity, Planarity, Scattering, and Verticality) were evaluated in terms of their contribution through class separability (F-value) and an ablation study. The results suggest that Scattering and Verticality are the dominant features contributing to classification performance.