All 30 cited studies

Grouped by topic · citation metadata verified via Crossref & arXiv · each line notes what the study contributes

Numbers match the small blue citation marks on the topic pages. Where a study is the source of a claim that failed fact-checking, that is noted — the paper is real; the over-general claim built on it is what was dropped.

Guardrails & barriers

Hou, Ai & Boudreau (2022). Network-Level Guardrail Extraction Based on 3D Local Features from Mobile LiDAR. J. Computing in Civil Engineering 36(6). ascelibrary.orgGeometric detection; 95.6% precision, 98–100% length coverage on two highways.

Vidal, Díaz-Vilariño, Arias & Balado (2020). Barrier and Guardrail Extraction and Classification from Point Clouds. ISPRS Archives XLIII-B5. isprs-archives.copernicus.orgRoad-corridor method; steel 0.91 vs concrete 0.70; source of the refuted "no road-path needed" claim.

Yue, Gouda & El-Basyouny (2021). Automatic Detection and Mapping of Highway Guardrails from Mobile LiDAR. IGARSS 2021. ieeexplore.ieee.orgTrajectory-corridor + clustering pipeline, label-free.

Jiang, He, Liu, Ai & Lang (2016). Corrugated beam guardrail detection based on mobile laser scanning. IEEE ITSC 2016. ieeexplore.ieee.orgScanline tracing of the W-beam's continuity.

Brkić, Miler, Ševrović & Medak (2022). Automatic Roadside Feature Detection from LiDAR Road Cross-Section Images. Sensors 22(15):5510. mdpi.comSide-on slices + image detector; type AP 0.98 concrete / 0.91 steel over 1,300 km.

Vertical markings — posts & signs

Gouda, Shalkamy, Li & El-Basyouny (2022). Fully Automated Pole Detection in Rural Highways from MLS. Transportation Research Record 2676(7). journals.sagepub.com98% precision/recall over 28 km; source of the refuted density-threshold claim.

El-Halawany & Lichti (2013). Detecting road poles from mobile terrestrial laser scanning. GIScience & Remote Sensing 50(6). tandfonline.comClassic 86% detection / 97% correctness baseline.

Ordóñez, Cabo & Sanz-Ablanedo (2017). Detection and Classification of Pole-Like Objects from MLS. Sensors 17(7):1465. mdpi.com91% detection; four-feature classifier 96% on type; source of the refuted "shape-ratio alone" claim.

Dong, Chen & Stachniss (2021). Online Range-Image-based Pole Extractor. arXiv 2108.08621. arxiv.orgOpen-source, runs on CPU (PRBonn/pole-localization).

Greenery / vegetation

Weinmann, Jutzi, Mallet & Weinmann (2017). Geometric Features and their Relevance for 3D Point Cloud Classification. ISPRS Annals IV-1/W1. bibliothek.kit.eduShape-statistics + random forest ~95%; names the most useful features.

Aditya, Lohani, Aryal & Winter (2023). Benchmarking Deep Learning for Urban Vegetation MLS Segmentation. arXiv 2306.10274 (IEEE TGRS). arxiv.orgNo single model wins across datasets; small 2.7M-param model recommended for limited GPUs.

Brodu & Lague (2011). Multi-scale dimensionality classification (CANUPO). arXiv 1107.0550 (ISPRS J.). arxiv.org>98% vegetation separation; the method behind the CANUPO / CloudCompare tool.

3D segmentation models & datasets (if the ML route is taken)

Wu et al. (2023). Point Transformer V3. arXiv 2312.10035 (CVPR 2024). arxiv.orgFast, memory-efficient 3D model; 16 GB feasibility here is untested (open question).

Robert, Raguet & Landrieu (2023). Efficient 3D Semantic Segmentation with Superpoint Transformer. arXiv 2306.08045 (ICCV). arxiv.orgVery small, fast model; attractive for limited hardware, to be tested.

Thomas et al. (2019). KPConv: Flexible and Deformable Convolution for Point Clouds. arXiv 1904.08889 (ICCV). arxiv.orgStrong outdoor baseline; vegetation-benchmark winner on Toronto-3D.

Hu et al. (2019). RandLA-Net: Efficient Segmentation of Large-Scale Point Clouds. arXiv 1911.11236 (CVPR 2020). arxiv.orgLightweight, handles ~1M points/pass on modest hardware.

Tan et al. (2020). Toronto-3D: Large-scale MLS Dataset for Roadway Segmentation. arXiv 2003.08284 (CVPRW). arxiv.orgPublic dataset with Road/Natural/Pole/Fence classes matching our targets.

Cortinhal, Tzelepis & Aksoy (2020). SalsaNext. arXiv 2003.03653 (ISVC). arxiv.orgTemplate for the "project-to-2D then segment" pattern.

Balado et al. (2019). Road Environment Semantic Segmentation with Deep Learning from MLS. Sensors 19(16):3466. mdpi.comDeep-learning segmentation of the road scene.

Jiang, Elmasry, Lim & El-Basyouny (2025). DL Semantic Segmentation of Multilane Rural Highways. Can. J. Civ. Eng. 52(8). cdnsciencepub.comRare highway result with an explicit guardrail class (~86% mIoU).

Label-efficient learning & hybrid 2D+3D

Liu et al. (2022). LESS: Label-Efficient Semantic Segmentation for LiDAR. arXiv 2210.08064 (ECCV). arxiv.orgNear-full accuracy from a fraction of labels.

Bayar et al. (2023). Point Cloud Segmentation via Transfer Learning with RandLA-Net. arXiv 2312.11880. arxiv.orgPretrain-then-fine-tune as a label-saving strategy.

Yarovoi & Valenta (2025). Data-Efficient Point Cloud Segmentation for Unimproved Roads. arXiv 2508.20135. arxiv.orgLabel-efficiency in a road (not urban) setting.

Yin et al. (2023). Weakly Supervised Segmentation of Large-Scale Industrial Point Clouds. Automation in Construction 148. sciencedirect.comWeak supervision replacing dense labels for infrastructure.

Fang et al. (2022). Joint Point-Cloud + Multi-View Network for Roadside Object Classification. ISPRS J. Photogramm. RS 193. sciencedirect.comPrecedent for combining rendered views with 3D points.

Ding et al. (2024). Scan-to-BIM for As-built Roads. arXiv 2406.12404. arxiv.orgReconstructs road assets (incl. guardrails) as geometry; 1.46 cm error.

Asphalt / road-edge extraction

Xu, Wang & Zheng (2016). Road Curb Extraction from Mobile LiDAR Point Clouds. arXiv 1610.04673 (IEEE TGRS). arxiv.orgWhy single height-step methods fail; multi-signal + continuity refinement.

Zhao et al. (2024). CurbNet: Curb Detection via LiDAR Segmentation. arXiv 2403.16794. arxiv.orgML option; useful class-imbalance handling if we go that route.

Wang, Ibrahim, Mansoor et al. (2025). Automated Road Extraction and Centreline Fitting in LiDAR. arXiv 2502.07486. arxiv.orgReference accuracy (67→73%); template for the smoothing/vectorization stage.

Schwab & Kolbe (2026). Radiometric Fingerprinting of Surfaces using MLS. arXiv 2603.11252. arxiv.orgWhy laser brightness must be used locally, never as a global threshold.