Finding guardrails, greenery & road edges in laser scans

Prepared for project leadership · plain-language version · four new deliverables assessed against the published state of the art

We already turn highway laser scans into clean CAD line drawings of the road markings. The client now wants four more things pulled out of the same scans. This briefing says, for each one, which method to build, how accurate it is in the literature, and how much work it is — backed by 30 cited studies.

Guardrailsposition & type (steel / concrete / cable)

Greenerytrees & bushes along the road

Vertical markingsdelineator posts & traffic signs

Asphalt edgewhere the pavement ends

▶ Listen: 22-minute audio explainer
A two-host walkthrough of this research — MP3 on Google Drive (ioLabs)

The one-sentence answer: for three of the four (guardrails, posts/signs, and the asphalt edge) the best route is classic geometry — rules written by hand that lean on the road path we already know — with no training data and no expensive graphics card. Only greenery genuinely benefits from machine learning, and even there a lightweight approach comes first.

Jump to a topic — each page is a short, self-contained read:

01Guardrails →Corridor geometry + a surprising "look along, don't look down" trick for reading the type. 02 · 03Greenery & posts →The one class that needs learning, and the pole-detection recipe that's basically solved. 04Asphalt edge →Trace the boundary of a surface we already extract — no curb needed. REFSources →All 30 studies, grouped and linked, with what each one proves.

Recommendation at a glance

Two words that matter throughout this briefing:

Geometry = we describe the shape in plain rules ("a guardrail is a long thin thing, about rail height, running parallel to the road"). No examples to collect, runs on a normal laptop CPU.
Machine learning (ML) = we show a network thousands of hand-labelled examples until it learns the shape itself. Powerful, but needs the labelled examples (which we'd have to make) and a graphics card.

What to detect	Recommended method	Backup / refinement	Confidence	Build effort
Guardrails	Geometry: search a narrow band along the road, find the long rail-height shape^1,2	Render side-on "slices" & classify steel vs concrete⁵	High	Low
Vertical markings (posts, signs)	Geometry: find thin, upright, tall clusters⁶	Tiny classifier for sign vs post vs sapling⁸	High	Low
Greenery (trees, bushes)	Measure local "shape statistics", feed a light classifier^10,12	Small 3D neural net if that's not enough¹¹	Medium	Medium
Asphalt edge	Trace the outline of the road surface we already extract, then refine²⁷	Texture/roughness profiles across the road³⁰	High	Low–Med

The accuracy, in plain bars

These are the headline results reported in the literature on real highway data. Higher is better; think of each as "how often it gets it right."

Guardrails found (precision)¹96%

Posts / signs found⁶98%

Steel-vs-concrete type (side-on)⁵91–98%

Greenery segmented¹⁰~95%

Pole type (tree/sign/lamp)⁸96%

Green = solid, high-confidence result. Amber = promising but from a single study or a different city, so treat as "very likely" until we test on our own scans. See each topic page for the caveats.

How much can you trust this?

The research was run twice, independently, and then fact-checked. The two runs never saw each other's work and reached the same recommendations.

independent research runs (cross-checked)

peer-reviewed sources cited

claims confirmed by fact-checkers

plausible-but-wrong claims caught & dropped

Every important claim was handed to three "checker" passes whose only job was to try to disprove it by reading the actual paper. A claim only survived if they failed to knock it down. Four claims that sounded right did not survive — those are the traps we now know to avoid, flagged in red on the topic pages.

Three takeaways worth remembering

Use what we already built. We know the road's path and we already extract the road surface. Between them, those two assets make most of this problem easy — before any new model is trained.
Match the viewing angle to the object. Flat things (paint, pavement edge) are best seen from above. Standing things (rails, posts, signs) are best seen side-on. This one idea reshapes how we render images from the scans — see the guardrails page.
Machine learning only where it earns its place. Here that's greenery, and nowhere else among the four.

Mini-glossary — the handful of terms used in this briefing

Point cloud — the raw output of the laser scan: millions of 3D dots, each with a position, a brightness value, and a colour.
MLS (mobile laser scanning) — the scanner is mounted on a vehicle driving the highway, so we also know the exact path it drove (the "trajectory" or road axis).
Segmentation — labelling every dot as "road", "guardrail", "tree", etc. This is what the ML models do.
Top-down vs cross-section — two ways to flatten the 3D dots into a 2D picture: looking straight down (like a map) or looking along the road at a thin slice (like an engineer's section drawing).
CAD polyline — the clean vector line we ultimately hand to the surveyors/Autodesk, the same format we already produce for road markings.

Produced by the KIRIOLL point-cloud team with an AI-assisted literature review (Claude Code deep-research workflow with 3-vote adversarial fact-checking, plus an independent Codex cross-check). Full source notes live in the project vault. Every number on these pages is traceable to a cited study on the sources page.