A 3D rail point cloud registration method based on the combination of PCA and GICP
Occlusion and missing surfaces complicate the alignment of LiDAR rail scans with standard models. This study proposes a rail-geometry-informed initialization strategy followed by generalized iterative closest point (GICP) refinement. A pose prior resolves PCA orientation ambiguity using a Frobenius-norm criterion, while a high-curvature rail-head region provides a geometric translation reference. Six methods were evaluated under identical preliminary alignment on 120 measured 0.5 m rail segments from an operational 60 kg/m rail line. The proposed pipeline achieved a whole-cloud point-to-surface RMSE of 1.848 ± 1.103 mm, compared with 7.831 ± 1.108 mm for GICP. The paired median improvement was 6.469 mm (95% confidence interval: 5.637–6.962 mm). The mean runtime increased from 0.41 to 2.52 s per segment. Additional experiments examined scanning distances of 0–10 m and the sensitivity to two initialization parameters. The results demonstrate the value of the proposed pipeline for scan-to-model rail alignment. It should be noted that generalization beyond the tested rail geometry and acquisition conditions remains unverified.
Authors
- Xinpeng Shi (ORCID: https://orcid.org/0009-0005-7896-6950)
- Xilong He (ORCID: https://orcid.org/0009-0008-2001-1942)
- Shengfang Zhang
- yan qin
- Ziguang Wang
- Siman Li (ORCID: https://orcid.org/0009-0008-2255-7682)
- Yu Liu
Institutions
- Dalian University of Technology (CN)
- Dalian Jiaotong University (CN)
Publication Details
- Journal
- Engineering Research Express
- Published
- 2026-11-18
- DOI
- https://doi.org/10.1088/2631-8695/aec412
- Primary Topic
- 3D Surveying and Cultural Heritage
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- National Natural Science Foundation of China