TunnelSDF-FilterNet: A Domain-Aware Neural Framework for High-Fidelity 3D Reconstruction of Railway Tunnels

Abstract

Accurate 3D reconstruction of railway tunnels is crucial for infrastructure maintenance, safety assessment, and digital twin development. However, existing methods often fail in real-world scenarios due to sensor noise and spurious geometry generation in non-structural ground regions—particularly caused by rails, sleepers, and ballast. To address these domain-specific challenges, we propose TunnelSDF-FilterNet, a domain-aware neural signed distance function (Neural SDF) framework explicitly designed for high-fidelity railway tunnel reconstruction. Our approach introduces three key innovations: (1) a robust preprocessing pipeline that integrates DBSCAN–RANSAC clustering with normal and height-based geometric filtering to automatically remove ground artifacts without manual intervention; (2) a tunnel-tailored Neural SDF training objective featuring a novel ground suppression loss to constrain reconstruction within the tunnel envelope and an edge-aware regularization term to preserve fine structural details such as segment joints; and (3) a NeuralPull-based surface refinement strategy applied during inference to achieve sub-voxel precision in surface extraction. Extensive experiments on real-world tunnel datasets demonstrate that TunnelSDF-FilterNet significantly outperforms state-of-the-art methods in both qualitative fidelity and quantitative metrics—including lower Chamfer Distance (CD), higher Normal Consistency (NC), reduced Ground Inclusion Rate (GIR), and improved F-Scores. The proposed framework delivers operationally deployable, high-precision 3D tunnel models, offering a practical AI-driven solution for intelligent railway infrastructure management and digital twin construction in modern railway systems.

Keywords

3D reconstruction / Railway tunnel / Point cloud / Neural SDF

References

[1] Zhang N, Peng L, He Y, Ye C. Flexible probe with array tunneling magnetoresistance sensors for curved structure inspection. IEEE Trans Instrum Meas, 2022, 71: 1-9

[2] Cao Y, He S, Wang F, Su S, Sun Y. A large-model-enhanced method for rail surface defect detection in heavy-haul railway. IEEE Trans Intell Transp Syst, 2025

[3] Peng X, Lin P, Xia Q, Yu L, Wang M. A new method for recognizing discontinuities from 3D point clouds in tunnel construction environments. Tunn Undergr Space Technol, 2024, 152 Article ID: 105955

[4] Ren Z, Liu Y, Zhao D, He Y, Zhang J. A deep learning-driven framework for real-time recognition and quantification of subway tunnel surface defects using high-resolution imaging. Urban Rail Transit, 2025

[5] Han H, Cai X, Gao L. A condition diagnosis method for subway track structures employing distributed optical fiber sensing. Sci Rep, 2025, 15(1): 28993

[6] Dong W, Xia X, Huang H, Zhong M, Hu Y, and Yang T (2025) Study on hydroaccumulation patterns in V-shaped slope tunnels of subways during floods. Urban Rail Transit, pp. 1–18

[7] Chen H, Li C, and Lee G. H, (2023) Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance, rXiv preprint arXiv. 00846

[8] Sun H, Xu Z, Yao L, Zhong R, Du L, Wu H. Tunnel monitoring and measuring system using mobile laser scanning: design and deployment. Remote Sens, 2020, 12(4): 730

[9] Cao Y, Wang F, and Su S, (2025) Bi-level sparsity augmented design method for selection of tractive locations of railway turnout. 275: 127123

[10] Wang P-S, Liu Y, Guo Y-X, Sun C-Y, Tong X. O–CNN: octree-based convolutional neural networks for 3d shape analysis. ACM Trans Graph, 2017, 36(4): 1-11

[11] Zheng L, Wang R, Wang S, Liu X, and Guo S, Point cloud plane fitting based on RANSAC and robust eigenvalue method. pp. 1368–1372

[12] Park J. J, Florence P, Straub J, Newcombe R, and Lovegrove S, Deepsdf: Learning continuous signed distance functions for shape representation. pp. 165–174

[13] Lu Y, Sun H, Zhang Z, Zhong R, Ren X, Sun M. End-to-end 3D reconstruction of subway tunnel scenes via local feature aggregation. ISPRS J Photogramm Remote Sens, 2025, 230: 563-580

[14] Yan X, Zheng C, Li Z, Wang S, and Cui S, Pointasnl: Robust point clouds processing using nonlocal neural networks with adaptive sampling. pp. 5589–5598

[15] Gropp A, Yariv L, Haim N, Atzmon M, and Lipman Y, (2020) Implicit geometric regularization for learning shapes, arXiv preprint arXiv. 10099

[16] Chen C, Liu Y S, Han Z (2023) Gridpull: towards scalability in learning implicit representations from 3d point clouds[C]//Proceedings of the ieee/cvf international conference on computer vision.: 18322–18334

[17] Ma B, Liu Y S, Zwicker M, et al. (2022) Surface reconstruction from point clouds by learning predictive context priors[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 6326–6337

[18] Li S, Gao G, Liu Y, et al. (2024) Implicit filtering for learning neural signed distance functions from 3d point clouds[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 234–251

[19] Wu C, Li X, Jiang R, Guo Y, Wang J, Yang Z. Graph‐based deep learning model for knowledge base completion in constraint management of construction projects. Comput Aided Civ Infrastruct Eng, 2023, 38(6): 702-719

[20] Tan L, Cao Y, Wang F, Tang T, Wang X, Li Q. A three-stage prediction method for track displacement during shield tunneling. Urban Rail Transit, 2023, 9(3): 205-220

[21] Young S. I, Dalca A. V, Ferrante E, Golland P, Metzler C. A, Fischl B, and Iglesias J. E, (2023) Supervision by denoising, IEEE transactions on pattern analysis and machine intelligence

[22] Li J, Song C, Lin J. Research on 3D reconstruction method of tunnel surface based on point cloud data. Urban Geotech. Investig. Surv, 2021, 6: 90-94

[23] Hai-ming GE, Wei Z, Xiao-long W, et al.. Optimization of 3D reconstruction method for urban cable tunnel based on SfM method[J]. Journal of Graphics, 2023, 44(3): 540

[24] Xu H, Zhou LS, Qiao YF. Tunnel 3D reconstruction technology and application research based on Kinect sensor. Tunnel Construction, 2018, 38(S2): 202-208

[25] Xu B, Wang Y, Hao W, Yin S, Wang W, and Li Y (2024) Neural-pull based on edge enhancement for 3D surface reconstruction,”Infrared and Laser Engineering

[26] Nezhadarya E, Taghavi E, Razani R, Liu B, and Luo J, Adaptive hierarchical down-sampling for point cloud classification. pp. 12956–12964

[27] Ouyang Q, Lin Y, Zhang X, Fan Y, Yang W, Huang T. Application of 3D reconstruction technology based on an improved MC algorithm in a shotcreting robot. Appl Opt, 2022, 61(29): 8649-8656

[28] Ram P, and Sinha K, Revisiting kd-tree for nearest neighbor search. pp. 1378–1388

[29] Qiu Z, Ma Y, Fan F, Huang J, Wu L, Du Y. Improved DBSCAN for infrared cluster small target detection. IEEE Geosci Remote Sens Lett, 2023, 20: 1-5

[30] Mescheder L, Oechsle M, Niemeyer M, Nowozin S, and Geiger A, Occupancy networks: learning 3d reconstruction in function space. pp. 4460–4470

[31] Li S, Gao G, Liu Y, Gu M, and Liu Y.-S, Implicit filtering for learning neural signed distance functions from 3d point clouds. pp. 234–251

[32] Ma B, Han Z, Liu Y.-S, and Zwicker M, (2020) Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces, arXiv preprint arXiv:.13495

[33] Tian D, Ochimizu H, Feng C, Cohen R, and Vetro A, Geometric distortion metrics for point cloud compression. pp. 3460–3464.

[34] Zhou L, Wu G, Zuo Y, Chen X, Hu H. A comprehensive review of vision-based 3d reconstruction methods. Sensors Basel, 2024, 24(7): 2314

RIGHTS & PERMISSIONS

The Author(s)