What BLE beacon tracking accuracy means on a crowded exhibition floor

In the supplier demo the dot glides along the aisle, turns into a stand, pauses for ninety seconds and leaves. It is convincing, and it was recorded in an empty hall with one person walking a rehearsed route.

On day two of a live show the same dot spends four minutes inside a competitor's stand while the person carrying the badge is standing in the aisle outside. Nobody has done anything wrong. BLE beacon tracking accuracy is set by a physical measurement that behaves badly in a room full of people, and the numbers on the datasheet were measured in a room that had none.

What the receiver actually measures

A Bluetooth Low Energy device advertises a short packet. A receiver hears it and records the received signal strength indication, a number in dBm that describes how much power arrived. That is the whole measurement. Distance is inferred afterwards, by assuming the power fell off with range in a predictable way.

The assumption is the log-distance path loss model: received power drops by a fixed number of decibels for every doubling of distance, with the rate set by an exponent that depends on the environment. In free space the exponent is 2. In a hall with bodies, carpet, trussing, double-decker stands and metal, it is whatever the hall makes it, and it is not constant across the floor.

Two things follow. Any distance a beacon system reports is the output of a model, with the model's assumptions inside it. And the conversion is logarithmic, so the error behaves in a way that surprises people who are used to linear instruments.

Turning signal strength into distance, and where the arithmetic gives out

Ramirez, Huang, Liao, Lin, Lin and Liang published a careful measurement study in Sensors in 2021, working through BLE RSSI for indoor positioning. Among their readings, signal strength runs at roughly minus 56 dBm at 0.2 metres and roughly minus 86 dBm at 10 metres.

Fit the path loss model to those two points and you can see the sensitivity directly. The distance ratio between them is 10 divided by 0.2, which is 50, and the base ten logarithm of 50 is 1.699. The signal fell 30 dB across that span, so ten times the path loss exponent is 30 divided by 1.699, which is 17.7. The exponent is about 1.77, a little below free space, which is what you would expect from a corridor with reflective surfaces.

Now invert it, because inverting it is what the positioning engine does. An error of E decibels in the reading corresponds to a distance ratio of ten raised to the power of E over 17.7.

A 3 dB error gives ten to the power of 0.170, which is a factor of 1.48. A 6 dB error gives ten to the power of 0.340, which is a factor of 2.19. A 10 dB error gives a factor of 3.68.

Put a real distance in front of that. A person genuinely 3 metres from a receiver, read 6 dB weak, is placed at 3 times 2.19, which is 6.6 metres. Read 6 dB strong, the same person is placed at 3 divided by 2.19, which is 1.4 metres. Six decibels is a small amount of interference. It moves the estimate across five metres of floor.

That asymmetry is worth holding on to as well. The same decibel error produces a much bigger absolute displacement outward than inward, so RSSI-derived positions have a long tail towards the far side of the hall.

Why does a full hall make it worse?

Because everything that changes between the demo and the show pushes on the same term.

Human bodies absorb energy at 2.4 GHz, so the attenuation between a badge and a receiver depends on how many people are standing between them, which changes minute by minute and peaks exactly when you most want the data. Stand builds, trussing and hanging banners add reflections, and a reflected packet arrived by a longer path and reads weaker than the direct line would.

The radio environment moves too. García-Paterna, Martínez-Sala and Sánchez-Aarnoutse, also writing in Sensors in 2021, ran a room-level BLE system in two settings and noted that in a university basement the radio spectrum was much more variable than in a house, because several wifi access points were broadcasting in the 2.4 GHz band and adapting their transmission power automatically. An exhibition hall is that condition with a few hundred more radios in it, including every exhibitor's own kit.

Their results also show something people rarely account for, which is that the receiver is part of the measurement. Running the same beacons and the same algorithm, a laptop scanned better than a Raspberry Pi in every configuration they tested. Whatever hardware your badge or your gateway uses is not interchangeable with whatever the pilot used.

What the published accuracy numbers actually cover

Read any accuracy claim as a claim about a specific geometry.

Ramirez and colleagues state the two cases plainly. Positioning could reach 10 centimetres when the beacons and scanners were on the same horizontal plane in a less noisy environment. The same work reports that accuracy dropped to metre scale once measurements were executed in a three-dimensional configuration and a complex environment. Both statements are true of the same equipment. A show floor, with receivers in the roof structure and badges at chest height among a crowd, is the second condition every time.

Centimetre figures do exist for Bluetooth outside the RSSI approach. The Bluetooth Special Interest Group, in its current published documentation on direction finding, describes the feature as supporting positioning systems that can achieve "centimeter-level location accuracy", and that feature works by measuring the angle a signal arrives from using an antenna array, which is a different measurement with different hardware on both ends. If a proposal quotes centimetres, the first question is whether it is selling angle of arrival or selling signal strength, because the two share a logo and nothing else.

Time of flight is a third family again, with its own cost structure, and what ultra-wideband buys and what it charges for it is C21's subject.

What accuracy question is worth asking a supplier?

Ask for the zone-level hit rate on a floor with people on it, at the zone size you actually intend to report.

That framing has published numbers behind it. García-Paterna and colleagues measured room-level classification rather than metre-level position, and the results are much better behaved. In a house divided into 10 rooms with 6 beacons, a laptop was correct 97.6 per cent of the time and a Raspberry Pi 87.7 per cent. Cut the beacons to 3 and the same two fell to 85.6 and 75.8 per cent. In a university basement divided into 16 zones with 10 beacons, the figures were 92.7 and 88.0 per cent, falling to 81.5 and 79.6 per cent with 4 beacons. They concluded that three beacons is the minimum for any usable performance in any scenario.

Two things transfer to a hall. Zone accuracy degrades gracefully as you thin the infrastructure, while metre accuracy falls off a cliff. And the zone has to be large relative to the error before any of it works.

Run that against a stand. Exhibition floors are usually laid out on a 3 metre module, so naming the correct stand requires the estimate to be right to better than 1.5 metres along the aisle, since that is where the boundary sits. At the metre-scale error the published work reports for cluttered three-dimensional spaces, that is close to a coin flip between a stand and its neighbour. Aggregate a coin flip over 4,000 attendees and you get a stand-level report that looks authoritative and is half noise.

Zones the size of an aisle run, ten to twenty metres across, put the error comfortably inside the zone, which is why paths held as ordered sequences of zones in C13 survive this and coordinate traces do not.

Where this stops

Accuracy is the second problem. The first is coverage, and it is larger.

A BLE system sees a badge when the badge advertises, when a receiver is in range, and when that receiver is listening on the channel the packet went out on. Miss any of the three and the person is simply absent from the data for that interval, with no error bar to warn you. Absence looks identical to a quiet zone, and no amount of accuracy work distinguishes them.

The second limit is that error is not uniform across the hall, so it is not something you can correct with one global adjustment. A zone under a mezzanine, a zone with a receiver behind a truss and a zone in open concourse have different exponents, different variances and different biases. A single quoted accuracy figure for a whole hall is an average over conditions that differ by more than the figure itself.

The third is that all of it changes on build day. Receivers get moved for rigging reasons, an exhibitor puts a two storey stand where the survey assumed open floor, and the calibration from the walkthrough is stale before the doors open. Anything you calibrate before build should be re-checked after it.

None of that makes the technology useless. It makes it a zone-level instrument, and picking the right instrument for the decision is the sensor selection question in C24, which sits at the front of any attendee analytics build.

This week, take one receiver, stand at a known point 3 metres from it with the hall empty, and log RSSI for two minutes. Then do the same thing at the same point during the busiest hour of the show. Subtract the two medians. Whatever that difference is in decibels, divide it by 17.7 and raise ten to that power, and you have the factor by which your own hall moves a distance estimate between an empty morning and a full afternoon.