When UWB positioning for events earns its cost over cheaper sensing

A quote lands for UWB positioning for events, covering one hall of a four-hall show, and the number has five figures in it before anyone has been tagged. The show director asks the only sensible question. What does this tell us that the door counters and the session scanners do not?

That question has a real answer, and it is narrower than the supplier's deck suggests.

What the technology actually measures

Ultra-wideband ranging times a radio pulse between a tag and a fixed anchor and converts that time into a distance. Three or four anchors with a clean path to the tag give a position by trilateration. The IEEE published 802.15.4z in 2020 as an amendment covering enhanced ultra-wideband physical layers and the ranging techniques that go with them, which is why phone-grade UWB and industrial UWB now behave broadly the same way.

The important word is path. Time of flight assumes the pulse travelled in a straight line. When it did, the arithmetic is very good. When it reflected off a truss or passed through two people first, the measured time is too long and the position is wrong by whatever that detour cost.

That single assumption is why the technology behaves so differently on a datasheet and on a Tuesday morning. Everything else in this post follows from it.

What does the accuracy claim mean once the hall fills up?

Every UWB datasheet quotes a number under ten centimetres. That number is measured in an empty room.

Liu, Lin, Wang and Kong, writing in Sensors in 2022, put the distinction plainly: ultra-wideband gives centimetre-level positioning accuracy in line-of-sight conditions, and walls and other obstacles introduce non-line-of-sight conditions that deteriorate that accuracy to the metre level. Their own test areas make the size of the drop visible. In one obstructed area with only two anchors holding line of sight, the root mean square error was 0.853 metres before correction, falling to 0.244 metres once they applied a delay model and to 0.132 metres with weighted least squares on top.

Read that as an operational fact about your own hall. A show floor is a machine for producing non-line-of-sight conditions. Fifteen thousand bodies, a hall full of two-storey stands, metal rigging, and a badge worn flat against a chest so that half the sky is blocked by the person wearing it. You spend some of the accuracy budget on physics before you spend any of it on hardware.

Volpi, Tebaldi, Matrella, Montanari and Bottani, also in Sensors, reported in 2023 on a real industrial deployment built on commercial UWB modules. Their static tests gave an overall average error always below one metre, with 0.16 metres on one axis and 0.31 metres on the other. That is a warehouse with racking, which is a friendlier environment than a show floor at 11:00 on day two.

The honest planning figure for a busy hall therefore sits somewhere between a few tens of centimetres and about a metre, depending on how much of your floor holds line of sight. Anyone quoting you ten centimetres for a live show is reading off a specification written for an empty room.

The anchor count is the whole capital budget

Here is where the money actually goes, and it is arithmetic you can do before you take the meeting.

Volpi and colleagues covered a 41 by 13 metre area, 533 square metres, with six anchors. That is one anchor per 89 square metres. Their other configurations ran four to six anchors over areas up to 588 square metres, so call the observed range one anchor per 90 to 150 square metres in an environment with obstructions.

Take a single hall of 4,000 square metres. At one anchor per 200 square metres, sparser than any of those deployments, you need 20 anchors. At the density Volpi and colleagues actually used, 4,000 divided by 89, you need 45. Those two numbers buy different things. Twenty anchors will resolve which aisle somebody stood in. Forty-five gets you close to resolving which stand.

Each anchor needs a surveyed position, a power source, and a mounting point that survives the build. On a show floor that means either the ceiling grid, which is a rigging job with a rigging cost, or truss and stanchions, which carry a build and derig cost every edition. The recurring labour is usually the part that surprises people, because the hardware is bought once and the hanging is bought annually.

The tag is the other half of the bill

Anchors are fixed. Tags scale with the audience, and that is where a portfolio-level decision gets made.

Volpi and colleagues priced 100 asset tags at about 2,000 euro, roughly 20 euro each. Take a show with 9,840 arrivals across three days and suppose you tag a sample, say the 850 buyers you most want to trace. At 20 euro that is 17,000 euro of tags, before the anchors, before the rigging, before the software.

Tag the whole audience instead and 9,840 tags at the same unit price is 196,800 euro. Nobody does that, which is worth saying out loud, because it means UWB at a show is almost always a sampled measurement. Once it is a sample, the sampling frame becomes the dominant source of error. A 0.3 metre position on a badge given only to pre-registered VIPs tells you a great deal about pre-registered VIPs and nothing about anybody else.

The cheaper instruments fail in a different place. A portal read tells you that a badge crossed a line at a time, and it carries no coordinate at all. Probe request counting has its own well-documented breakdown, because phones stopped presenting a stable identifier, and the wifi footfall counting limits that follow from address randomisation are a separate argument with a separate fix.

When is UWB worth it over cheaper sensing?

The test I would apply is whether any decision changes when the position error falls from three metres to thirty centimetres.

If the question is which zones were busy on Tuesday afternoon, zone counts from portals answer it and UWB is an expensive way to reach the same answer. If the question is how long a buyer stood at one 3 by 3 metre stand as against the one beside it, portals cannot answer it at all, and BLE proximity will confuse the two, because a two to three metre error band covers both stands and the aisle between them.

Two decisions genuinely need the resolution. The first is per-stand attribution inside a dense zone, where you are telling an exhibitor something about their own booth instead of about their aisle. The second is queue and bottleneck geometry, where knowing that the pinch sits four metres inside the door changes where you put a steward.

Everything else on a show floor is a zone question, and zone questions are answered by portals and readers at a fraction of the cost. Choosing between those instruments across a whole hall is a broader exercise, and the trade-offs across people counting sensors run wider than any one technology.

What arrives in the warehouse, and what it costs to store

One practical point that rarely reaches the business case. A UWB system at a 10 Hz update rate produces ten position rows per tag per second. Eight hundred and fifty tags, worn for six active hours a day, is 850 times 10 times 21,600 seconds, which is 183.6 million rows a day and about 551 million across a three-day show.

You will not report on 551 million rows, and you should not try. The useful pipeline downsamples to a zone assignment per tag per second or per five seconds, keeps the raw stream only long enough to re-derive that assignment if the zone polygons change, then throws it away. Deciding that retention rule before the show rather than after it is the difference between a warehouse bill you planned for and one you discover. This is a general point about how on-site data reaches attendee analytics, and UWB is simply the noisiest source of it.

Where this stops

UWB gives you a position. It says nothing about intention, and it is silent about the two things most likely to break your analysis.

The first is tag compliance. A tag in a lanyard flipped backwards, left on a chair over lunch, or handed to a colleague produces a perfect position for the wrong person. Nothing in the data announces that this happened. You can partially detect the extreme cases with implausible trajectories, a tag moving at four metres per second or sitting motionless for three hours in an aisle, but the mundane failure of a badge turned the wrong way just quietly loses read rate at the worst moment.

The second is that anchor geometry decides accuracy as much as anchor count does. Four anchors in a line give a good distance along that line and a poor one across it. Volpi and colleagues found notably different errors on their two axes, 0.16 against 0.31 metres, from exactly this effect. Somebody has to survey the anchor positions and record them, and that record has to survive being rebuilt next year in a hall with a different floorplan. If the survey is half a metre out, every position downstream is half a metre out and nothing in the data will tell you.

The step worth taking this week costs nothing. Print your floorplan, mark every point where you would need to distinguish two neighbouring stands from each other, and count the square metres those points cover. If that area is a small part of one hall, price a partial deployment over it and leave the rest on portals. If the marked area comes out empty, you have your answer about UWB for this edition, and it took an afternoon instead of a procurement cycle.