Automated dimension estimation of bridge components using semantic segmentation and geometric fitting of point cloud data
Yu Chen, Chao Lin, Xianfeng Li, Shiori Kubo, Tatsuro Yamane, Pang‐jo Chun
Abstract
Accurate estimation of dimensions for bridge components is essential for maintaining infrastructure and ensuring safety. This study proposes an automated framework that integrates deep learning-based semantic segmentation with geometric analysis for bridge component dimension estimation from point cloud data. The framework comprises two main stages: (1) a semantic segmentation module based on the PointNet++ architecture with MSG, enhanced with the improved structure-oriented concept (SOC) for improved component recognition, and (2) a geometric processing pipeline that incorporates voxel-based downsampling, RANSAC plane fitting, and multi-stage filtering for robust dimension extraction. The improved SOC mechanism effectively utilizes the inherent spatial relationships of bridge components, significantly enhancing segmentation accuracy in complex structural contexts. The geometric processing pipeline employs an adaptive region-growing algorithm for component isolation, followed by principal component analysis for orientation alignment and dimension calculation. Validation experiments conducted on two bridge datasets of varying scales demonstrated the effectiveness of the framework, achieving an overall relative percentage error (RPE) of 5.05% and a strong correlation (Pearson r = 0.904) between predicted and actual measurements. The method consistently performed well across various components, across two datasets the framework achieved mean relative errors of 2.6 % (length) and 5.2 % (width) for smaller bridges and 6.9 % (length) and 5.5 % (width) for larger bridges, confirming both accuracy and scalability. Statistical Bland–Altman analysis showed minimal systematic bias, underscoring the method’s practical value for automated bridge inspection and digital-twin applications.