The question that stops a delivery, a construction job, or a fire appliance is nearly always the same one: how wide is the street. Not how long, not what it is called, not what class it was assigned. How much clear space is there between the things on either side.
It is a question that most map data answers badly. Road networks are stored as centrelines, which are one-dimensional objects with no width at all. Width, when it is present as an attribute, is usually missing, occasionally wrong, and rarely dated. The information exists in the world and mostly not in the map.
There are ways to get at it remotely, and they differ not just in accuracy but in what kind of claim they support. Getting that distinction right matters more than squeezing the last half metre out of any single method.
Why width is the number people actually want
Access is the first reason. A site is only usable if the vehicles that serve it can reach it. A distribution centre served by a road too narrow for an articulated lorry to turn into is not served. Fire service access requirements are expressed in metres of clear width and headroom, and a route that fails them fails regardless of what the map says about connectivity.
Passability is the second, and it is dynamic. A street with a carriageway nominally wide enough for two-way traffic becomes single-lane when cars park along one side and impassable when they park along both. The gap between built width and usable width is where most last-mile routing failures live, because routing engines consume the network graph, and the graph knows about connectivity, not clearance.
Capacity is the third. Lane count, which drives most traffic modelling, is a function of width. Where lane counts are unmapped, width is the best available proxy, and it is a much better one than road classification.
There is a fourth use that is less discussed: width signals the character of a place. Wide streets with deep setbacks and narrow streets with buildings hard against the kerb support different densities, retail patterns, and transport modes. Width is one of the few physical measurements that correlates with urban form without needing socioeconomic data.
Decide which width before measuring anything
Half the disagreements about street width are definitional rather than empirical. There are at least four distinct measurements that all get called width:
- Carriageway width: the paved surface intended for vehicles, kerb to kerb.
- Trafficable width: the clear space actually available after parking, street furniture, market stalls, and encroachment. Almost always narrower than the carriageway, and the number that determines whether a vehicle passes.
- Right of way width: the legal corridor, which includes footways, verges, and sometimes land beyond the visible street. Used in planning and largely irrelevant operationally.
- Building to building width: the gap between the facades on either side. Relevant to canyon effects, insolation, wind, and radio propagation, and generally much wider than the carriageway.
These can differ by a factor of three on the same street. A method that measures facade to facade and a method that measures kerb to kerb will disagree permanently and neither is wrong. Every width value should carry which definition it uses, and a pipeline that mixes definitions across sources will produce a distribution that means nothing.
Three approaches, three different kinds of claim
The first approach is a class-based prior. Take the road classification from the map, look up a typical width for that class in that country, and use it. A motorway gets a wide value, a residential street a narrow one, a track narrower still.
This is cheap, complete, and available everywhere the network is mapped. It is also not a measurement of the street in question. It is a statement about what streets of that class usually look like, applied to a street nobody looked at. Its error is not random: it is systematically wrong in exactly the places where classification is loose, which is most of the world outside the countries whose tagging conventions the classification scheme was designed around. A residential street in a dense informal settlement and a residential street in a low-density suburb carry the same tag and differ by five metres.
The second approach is geometry from mapped edges. Where the map contains the features that bound the street, width can be computed rather than assumed. Kerb lines, road area polygons, building footprints on both sides, or parcel boundaries all bound the corridor. The computation is a cross-section: sample points along the centreline, cast a perpendicular at each, find the first intersection on each side, and take the distance.
This is a genuine measurement, but of the map rather than of the street. Its accuracy is bounded by the positional accuracy of the bounding features, which for traced building footprints can be a metre or two per side, and it inherits every gap in the map. Where footprints are missing on one side, the perpendicular runs off into open ground and returns an absurd value. Where the map has a road polygon, this method is strong. Where it has only centrelines and buildings, it measures facade to facade and must be labelled accordingly.
The third approach is measurement from street level imagery with depth. Take a photograph from within the street, recover per-pixel depth using a monocular depth model or stereo, identify the surfaces that bound the corridor, and compute the distance between them in the recovered geometry.
This is the only one of the three that measures the actual street as it actually is, including the parked cars. It is also the one with the most ways to go wrong, and its errors are correlated with scene content rather than being independent noise.
A prior is not a measurement
The distinction is worth stating plainly because pipelines erase it constantly.
A prior encodes what is typical. Its error distribution is the variation across the population of streets in that class. It is right on average and it carries no information whatever about the specific street. Averaging a million class-based priors gives a good estimate of the total road area in a country. Using one to decide whether a lorry fits is not analysis, it is a guess wearing a number.
A measurement encodes what was observed. Its error distribution comes from the measurement process and is usually much tighter, and critically it is specific: it can be wrong, but it can be wrong about this street rather than about streets in general.
The failure that follows from confusing them is not that the prior is inaccurate. It is that a system fed a mixture of priors and measurements, all stored in the same numeric field, cannot distinguish the two afterwards. Every downstream consumer treats them identically. The route planner that refuses to send a vehicle down a measured three metre street will happily send it down a prior-derived six metre street that is actually two and a half.
The fix is structural, not statistical. Store the method alongside the value, in a field that consumers cannot avoid seeing, and make it impossible to retrieve a width without retrieving how it was obtained.
Cross-checking two independent methods
The strongest practical technique available is to compute width by two methods that fail differently, and compare.
When the geometric method and the imagery method agree to within a small margin, the confidence in the value should rise substantially, because the two methods have almost nothing in common. One depends on the accuracy of digitised footprints and the completeness of the map. The other depends on camera pose, depth estimation, and surface identification. There is no shared error mode that would make both wrong in the same direction by the same amount. Agreement is therefore evidence, not coincidence.
Disagreement is more interesting, because it is diagnostic rather than merely negative. The direction of disagreement narrows the cause considerably.
- Imagery reports narrower than geometry: usually obstruction. Parked vehicles, stalls, containers, or vegetation are occupying the corridor, and the imagery is measuring the trafficable width while the geometry is measuring the built width. This is not an error. It is the most operationally valuable observation the system can make, and it should be surfaced as an obstruction estimate rather than suppressed as a mismatch.
- Imagery reports wider than geometry: usually a map problem. A missing building on one side, a footprint drawn short, or the geometry method having latched onto a boundary wall rather than the facade.
- Both plausible but far apart with no consistent pattern: usually a pose or scale problem in the imagery pipeline, or a datum offset in the vector data.
A single method returns a number. Two methods return a number and a diagnosis, and the diagnosis is often worth more than the number.
Where the error comes from
For imagery-based measurement specifically, the error budget is dominated by a handful of things, and knowing which one is active determines whether a value is salvageable.
Parked vehicles are the largest and the most ambiguous. A car occupies roughly two metres of the corridor and its presence is transient. Whether it should be included depends entirely on the question: for a delivery at that hour it is real, for a capacity model it is noise, for a planning application it is irrelevant. The measurement is not wrong either way, but the interpretation depends on a decision nobody usually records.
Camera pose is the largest purely technical source. Recovering metric distance from an image requires knowing where the camera was, which way it was pointing, and its intrinsics. Errors in heading translate directly into errors in which surfaces the perpendicular hits. Errors in height above ground scale the whole reconstruction. Consumer devices report pose with accuracy that varies with environment, and urban canyons, which is where narrow streets are, are exactly where positioning degrades most.
Sloped ground breaks the common assumption of a flat ground plane. A camera on a cambered or hilly street measuring across the corridor is measuring along a surface that is not perpendicular to the assumed vertical, and the resulting width is biased. On steep terrain the bias is not subtle.
Obstructions and occlusion mean the true boundary is frequently not visible. A wall behind a hedge, a facade behind a delivery van, a kerb under a pile of building material. The measurement then captures the obstruction's face rather than the boundary, and it will do so consistently, which makes it look like a stable measurement rather than a biased one.
Lens distortion, rolling shutter on moving platforms, low light, and rain on the lens all contribute. They mostly degrade the depth estimate rather than biasing it, which is preferable, since noise that averages out across multiple observations is easier to handle than bias that does not.
Calibration against something of known size
Any depth-based pipeline needs an anchor, because monocular depth estimation is generally accurate about relative depth and unreliable about absolute scale. A scene can be reconstructed perfectly in shape and be uniformly ten percent too large.
The practical answer is to find something in the frame whose real dimensions are constrained. Standard road markings have specified widths in most jurisdictions. Traffic sign faces are made to standard sizes. Vehicle dimensions cluster tightly by class. Door heights, brick courses, kerb heights, and standard paving units all sit within narrow ranges. Recovering a scale factor from any of these and applying it to the scene converts a relative reconstruction into a metric one.
Three cautions apply. Standards differ by country, so a calibration table has to be regional or it will introduce a systematic national bias. The calibration object must be at a similar depth to the surfaces being measured, since applying a scale recovered from a near object to a far one propagates error. And calibration should be verified rather than assumed: where more than one candidate object is visible, the scale factors they imply should agree, and if they do not, the reconstruction has a problem that a single anchor would have hidden.
Where no calibration object is available, that is worth recording. A width derived without a metric anchor is a different quality of number from one derived with two agreeing anchors, and the difference should be visible in the output.
Reporting a width so it will not be over-trusted
A bare number invites the reader to assume it was measured and to assume it is current. Both assumptions are frequently false, and the reporting format is where that gets corrected or entrenched.
A width value should carry, at minimum:
- The definition used. Carriageway, trafficable, right of way, or facade to facade.
- The method. Class prior, vector geometry, imagery with depth, or a fusion, and if fusion, which inputs.
- An interval rather than a point where the method supports one. A street reported as between four and five metres is more useful than one reported as 4.37 metres, because the second implies a precision the method cannot deliver and the first tells the reader what decisions it can support.
- The observation date, which for imagery-derived values is the capture date of the photograph and not the processing date. Street imagery can be years old, and streets narrow over time as parking pressure and encroachment increase.
- Whether obstructions were present and whether they were included.
- The number of independent observations behind the value. One pass down a street in one season is a different claim from twelve passes across two years.
Rounding is part of honesty here. Reporting a class-derived prior to two decimal places is a false precision problem that costs nothing to avoid: quantise priors coarsely so they visibly differ in texture from measurements. When a value looks approximate, it gets treated as approximate.
The decision rule
Before a width value is used for anything with a physical consequence, it should be possible to answer whether the number describes this street or streets like it. That single question separates a measurement from a prior, and it is the only distinction that matters at the point of use.
If the answer is streets like it, the value is fine for aggregate analysis and unfit for dispatch. If the answer is this street, the next question is when, because a measurement from imagery captured four years ago describes a street that may since have gained a cycle lane, a row of bollards, or a permanent line of parked vans.
The number that causes damage is not the wrong one. It is the right one, measured well, reported without its method, and reused three years later by someone who assumed it was current.