C-DRONE GUIDE · 5 SEPTEMBER 2026

Level crossing safety diagnostic: what a drone survey adds (sight distance, longitudinal profile, traffic) and what it costs

Since 2021, a French municipality or département managing a road that crosses a railway must produce a road safety diagnostic for each level crossing concerned, in coordination with SNCF Réseau, and refresh it every five years. Of the roughly 15,000 level crossings on France's national rail network, more than two-thirds are crossed by a municipal road: most of that workload therefore falls on authorities with neither an in-house survey team nor an engineering department. Yet the core of the diagnostic — sight distances, road geometry at the crossing, actual traffic — is hard to measure with a tape and hard to document with ground photos. Here is exactly what the regulations require, what a drone survey lets you put hard numbers on in the diagnostic grid, what it does not replace, and what it costs.

Published on 5 September 2026, reviewed on 5 September 2026 — regulations in force as of September 2026.

What the law has required of road authorities since 2021

The French transport ministry counted 15,405 level crossings in 2017 on lines open to traffic on the national rail network, split 0.4% on national roads, 31.4% on departmental roads and 68.2% on municipal roads. More than a hundred accidents occur there every year: SNCF Réseau recorded 89 accidents, 20 deaths and 10 seriously injured in 2024, then 112 accidents, 22 deaths and 13 seriously injured in 2025. The infrastructure manager stresses that 98% of those accidents involve road users breaching the highway code — which says nothing about the layout itself, but points the diagnostic at a precise question: what behaviour does the site configuration induce?

A national safety programme has existed since 1997, defined by the State and the national level crossing body: it listed 437 priority crossings at its creation and 155 as of 7 November 2018 — a figure still quoted in a 2025 ministerial answer to the Senate. Around €60 million a year is spent on level crossing safety by public authorities, and heavy operations (removing a crossing through a grade-separated structure) are co-financed by the State, SNCF Réseau and the road authority, on a split negotiated project by project rather than set by a single rule.

Since article 125 of the December 2019 mobility orientation act, codified at article L. 1614-1 of the transport code, the obligation is no longer limited to crossings in the national programme: every road authority must produce and keep up to date, in coordination with the railway infrastructure manager, a road safety diagnostic of the level crossings where their networks intersect. Decree n° 2021-396 of 6 April 2021 (articles R. 1614-2 to R. 1614-6) and the order of 3 May 2021 set out the details: the diagnostic identifies risk factors when crossing, by day and by night, is valid for five years and must be updated sooner if the crossing or its surroundings change. The document has four parts: a sheet of the crossing's main characteristics, an attendance sheet, an assessment grid covering geometry, visibility and legibility conditions and crossing by active modes, and a recommendations grid with an indicative timetable. The prefect then forwards the documents to the body running the public diagnostics database, currently hosted by the Cerema.

The Cerema also supplies the method: its report Road safety diagnostics of level crossings — tools and support, published on 20 February 2023, offers an analysis grid to be filled in jointly during the site visit and updates the 2008 Sétra methodology. Finally, the classification and equipment of crossings remain governed by the order of 18 March 1991, still in force and amended notably in 2017: four categories, from the 1st (public crossing with full or half barriers) to the 4th (private crossing), by way of the 2nd — the public crossing without barriers, marked by St Andrew's crosses, whose conditions of existence are entirely quantified.

Sight distances: calculated values, not a field impression

This is where the diagnostic turns technical. Annex I of the order of 18 March 1991 does not ask whether "you can see the train well": it imposes calculated distances. For an observer stopped 3.50 to 5 metres from the nearest rail, the required sight distance along the track is R1 = 0.8 F √(n + 5.6), and R2 = (3.4 + 0.7 n) F where the crossing is used by vehicles over 14 metres long or large herds — F being maximum train speed in km/h and n the number of tracks; both must remain below 600 metres. Two further distances, L1 and L2, apply to an observer in motion on the road and additionally depend on road speed V. In other words: on a double-track line at 100 km/h, the question is not whether you can see a train, but whether you can see it several hundred metres away from a point four metres from the rail.

Measuring that on site is tedious and rarely reproducible: you have to stand at the right point, sight along the track, estimate where the hedge, embankment, boundary wall, corner building or catenary mast cuts the line of sight, and start over for each direction of road travel and each direction of train approach. A drone photogrammetric survey changes the nature of the exercise: you rebuild a 3D model of the surroundings (terrain, vegetation, buildings, fences, platforms, street furniture) and then compute sight distances within the model, as many times as needed, with observation points placed exactly where the annex requires. That is the approach validated by a study from L. Iglesias, C. De Santos-Berbel, V. Pascual and M. Castro published in 2019 in Remote Sensing: a double-grid flight, a point cloud from Structure from Motion photogrammetry, a classification separating the terrain model from 3D objects, then a calculation of available sight distance in a GIS — the authors detected accident-prone locations caused by sight distance limitations, and measured how the 3D modelling parameters influenced the result (see the study on Google Scholar). Their test case was a tree-lined road — exactly the configuration that causes trouble around a rural level crossing.

One consequence is a deliverable that pays off beyond the diagnostic itself: a clearance simulation. By removing the offending hedge or thicket from the 3D model, you obtain the sight distance that would be restored, and therefore the quantified argument to put to the neighbouring landowner or to write into the recommendations grid, with proportionate works costs.

One honest caveat is needed here, and it comes from the literature. An experimental study by N. J. Ward and G. J. S. Wilde, published in 1996 in Safety Science, measured driver approach behaviour at an unprotected level crossing before and after lateral sight distances were improved: visual search times lengthened, but approach speeds tended to rise, so that the works produced no calculated net safety benefit (see the study on Google Scholar). Clearing sight lines is therefore necessary to comply with Annex I, but not sufficient on its own: the diagnostic must pair the geometric measurement with treatment of approach speed and signage legibility, or the gain dissipates into user behaviour.

Longitudinal profile and platform: the hump that strands an articulated lorry

This is the second hard point of the diagnostic, and the least well documented in local authority files. In its method, the Cerema explicitly addresses the case of an uneven longitudinal profile — a dip or a hump right at the crossing — liable to slow the crossing manoeuvre or even to strand a low-slung vehicle on the crossing. The mechanism is purely mechanical: the railway platform sets its own level, the road connects to it via two ramps, and the connection creates a high or low point that the overhang of an articulated lorry, a plant transporter or a bus cannot clear without grounding. Where the Cerema judges the problem structural, its recommendation is not road marking but heavy works: restore the longitudinal profile, correct the curves, widen the shoulders, rework the cross-section on the approaches.

But you first have to prove that the hump exists, and by how much. A drone photogrammetric survey produces a digital terrain model of the road and rail footprint, from which you then extract what no ground photo can give: the actual longitudinal profile of the carriageway over 50 to 100 metres either side of the track, a series of cross-sections, the gradients of the approach ramps, the usable width of the platform and shoulders, and the position of equipment. Our guide to drone photogrammetry deliverables sets out what those products technically contain and the accuracy you can expect. On that basis the check becomes geometric: you project the reference vehicle's envelope — a 16.50 m articulated lorry with its ground clearance — onto the surveyed profile, and see whether the overhang clears. That kind of quantified demonstration carries far more weight in a co-funding application than a photo of a lorry in trouble.

The same dataset then feeds the improvement scheme. A georeferenced orthophoto of the junction, backed by the point cloud, becomes the base plan on which the designer draws a relocated private access, a lengthened island, a corrected curve radius or a segregated pedestrian route — that last question appearing explicitly in the assessment grid, which covers crossing by active modes. Finally, the survey serves as a dated before/after record, reusable at the five-yearly update of the diagnostic. Where the issue is vegetation encroaching on the railway land, our guide to mapping vegetation alongside railway lines covers the railway side of the subject.

Road traffic, rail traffic and the traffic moment: measuring instead of estimating

The level crossing is the only piece of road infrastructure whose regulation rests on a crossed indicator: the traffic moment, the product of the number of trains by the number of road vehicles using the crossing on an average day. That indicator is not decorative: article 19 of the order of 18 March 1991 makes equipping a crossing with St Andrew's crosses plus Stop signs conditional on a traffic moment not exceeding 5,000 and on daily road traffic of at most 100 vehicles on average, the visibility condition being satisfied separately. A completed housing development, a rerouted itinerary, a new access to a business park, and the crossing can slip out of its category without anyone noticing — whereas rail traffic is known to the infrastructure manager down to the individual train.

Yet road traffic on these municipal roads is almost always estimated, rarely measured. A drone count removes the uncertainty in a single session: a camera hovering above the junction across representative time slots produces video whose analysis yields flow, directional split and heavy goods share — the last of these weighing most at a level crossing, between the grounding risk and the R2 sight distance triggered by vehicles over 14 metres. Our guide to drone traffic counting covers the method and its limits. Two points are worth stating up front: a count over a few hours is not equivalent to a permanent counter over a week, and you have to choose slots that capture the site's real peaks — the school run, farm access at harvest, quarry traffic.

The flight also lets you observe what statistics do not show and what the diagnostic grid is looking for: behaviour. Drivers weaving around half-barriers, vehicles queuing across the crossing when the junction downstream is saturated, pedestrians and cyclists taking desire lines outside the designated routes, cyclists steering to avoid skewed rails. Those dated, time-stamped observations feed the recommendations grid far more usefully than a declarative statement. They remain surface observations, however: they describe what users do, not why, and do not replace the night-time joint visit the diagnostic requires.

Flight framework, limits of the method and 2026 prices

Two authorisation chains overlap at a level crossing, and neither excuses the other. On the rail side, proximity to an operational track requires coordination with SNCF Réseau and compliance with the applicable clearance distance: our guide to flying a drone near a railway line covers the 30 m rule and the process to follow. The good news: since the diagnostic itself mandates coordination with the infrastructure manager, the contact is already identified — this is one of the rare cases where the administrative side of the overflight rests on a pre-existing regulatory obligation. On the road side, occupying the highway to take off or to set up an observation point falls to the road authority and, where a flight must take place over running lanes, to traffic policing: our guide to flying over a road in service covers the road permit and the temporary traffic order. On top of that sits the ordinary aviation framework (flight category, zones, height) under drone regulations.

Now the limits, and they are real. A drone survey does not sign the diagnostic: the document is a joint one, carries an attendance sheet, and presupposes an on-site visit bringing together the road authority and the infrastructure manager, by day and by night. The night-time component — perception of the flashing red signals, glare, legibility of advance signage in headlights — is poorly documented by photogrammetry and remains a matter of human observation. Nor does a 3D model capture moving or seasonal objects: the stack of bales at the field edge that masks the track two months a year, the vehicle parked in front of the St Andrew's cross, the hedge measured just after the winter cut. Finally, the accuracy of a sight distance calculation depends directly on model quality — cloud density, terrain/object classification, geodetic control: without GNSS-surveyed control points you get orders of magnitude, not defensible values.

Orders of magnitude observed in France in 2026, excluding VAT, for a survey intended to feed a level crossing diagnostic:

These amounts cover the flight, data processing and the report; they cover neither the diagnostic itself, which remains the road authority's and its consultant's job, nor the works supervision. Pooling is the main saving here: a département or a group of municipalities that has the ten or fifteen crossings on a single line surveyed in one go amortises travel, coordination with SNCF Réseau and the setup of the sight distance computation across the whole run, instead of paying for them crossing by crossing. Request a quote stating the list of level crossings concerned, their category and your diagnostic deadlines: that is what sizes the assignment.

Frequently asked questions

Request a free quote

Put it into practice

• Drone security & surveillance: rates and cities covered from €600
• Security & surveillance in Brive-la-Gaillarde Nouvelle-Aquitaine
• Security & surveillance in Sète Occitanie
• Security & surveillance in Bourg-en-Bresse Auvergne-Rhône-Alpes

Also worth reading

• Drone spraying in banana plantations against black Sigatoka: 2026 legal framework, method, price (French Antilles)
• SPANC and onsite wastewater systems: what a drone survey prepares — and what it never concludes
• Detect and avoid (DAA): how a drone operator proves it controls the mid-air collision risk