Guardrails

Detect position and type — steel W-beam, concrete barrier, or cable · High confidence Low effort

Recommendation: find guardrails with plain geometry (no training data, runs on a laptop CPU), then read the type from a side-on image slice. Both halves are proven on real highways.

Why this is easy: we already know where to look

A guardrail is a long, thin thing at roughly rail height that runs parallel to the road. Because the scanner truck recorded the exact path it drove, we can ignore everything except a narrow band a few metres either side of the road. In one study that single step threw away 96% of the points and the whole detector then ran in minutes on an ordinary laptop — no graphics card.²

✓ Confirmed by fact-checking Three separate, peer-reviewed methods find guardrails this way with no training data at all — using the road path, the rail's height, and its continuity along the road.^1,2,3,4

How accurate is it?

Guardrails found — precision¹96%

Steel W-beam rails²91%

Concrete barriers²~70%

Steel rails are the easiest of all — their wavy shape is distinctive. Concrete barriers are harder, simply because a concrete wall looks like other walls (noise fences, parapets). And the main thing that makes a rail go undetected is bushes growing over it — no amount of rule-tuning sees through a hedge.

The surprise: to read the type, look along the road — not down

The obvious plan was to reuse our top-down map images. But the one team that made image-based detection work at scale (1,300 km of highway) deliberately looked side-on instead — slicing the road like a loaf of bread and reading each slice like an engineer's section drawing.⁵

The same guardrail from two angles. From above it collapses to a faint stripe; side-on, its profile is unmistakable — which is exactly the shape that tells steel from concrete.

On those side-on slices, a standard image detector read the barrier type at 98% (concrete) and 91% (steel) — the best of every roadside object it tried.⁵ The two methods compose neatly: geometry finds where the rail is; a side-on slice tells us what it is. Drawing boxes on those slices is also about the easiest labelling job there is — and a colleague already prototyped this exact side-on rendering earlier this week.

✗ Two tempting claims that did NOT survive fact-checking

"You can skip the road path and it still works" — rejected (0 of 3 checkers). These methods lean heavily on the road path; it is essential, not optional.²
"Already proven at full production scale" — rejected (0 of 3). The best result was validated on two routes, not a whole network.¹

⚠ Genuine gap: cable barriers No solid published result exists for the third type — steel cables on posts. Thin cables give sparse laser returns. If our highways have them, we will need to build and test our own detector.

➜ Recommendation Build the geometric corridor detector first (label-free, CPU-only, ~96% precision). Add a side-on slice renderer and a light image detector for the steel/concrete/cable type. Treat cable barriers and dense-vegetation occlusion as the two items to validate on our own data.