Where lidar people counting accuracy holds up and where crowds break it

The quote says 99 per cent at the doorway and the show director signs it, because 99 per cent sounds like a solved problem. Three months later the door total for day two reads 12,700, the badge scans read 11,800, and nobody in the room can say which number is closer to the truth or where the difference came from.

Lidar people counting accuracy is a real and generally good property of the technology. The trouble is the shape of the claim. It is quoted per crossing, measured at a doorway, usually under conditions the supplier chose, and none of those three qualifications is written on the slide that ends up in the board pack.

What an overhead lidar counter actually does

The sensor fires laser pulses and times the returns, so what it holds is a set of distances. Mounted overhead at a doorway, it sees the floor at a known range and anything closer than the floor as an object.

From there the work is clustering. Filter out the static background, group the remaining returns into clusters, decide which clusters are people, track each one across frames and count it when it crosses a line. Direction comes from the track.

Lesani, Nateghinia and Miranda-Moreno set the whole pipeline out in Transportation Research Part C in 2020, in a paper running from page 20 to page 35 of volume 114. Their system used a sixteen channel unit taking distance measurements every 20 milliseconds, applied background removal, then ran two separate clustering routines they called LRC and HRC to produce single clusters and group clusters. They report that the system accurately counts more than 97 per cent of the pedestrians at the disaggregate level, with a false direction detection rate of 1.1 per cent.

The detail worth pausing on is that they needed two clustering routines. One general routine could not handle both an individual walking alone and four people arriving abreast, so the design has a low resolution stage and a high resolution stage. Groups are the hard case, and the published method treats them as a separate problem, which is a stronger statement about difficulty than any error bar.

What a 99 per cent figure leaves behind

Take the day above. Hall one recorded 12,700 arrivals.

At the quoted 99 per cent, the sensor misses 1 per cent, which is 127 people. At the published 97 per cent, it misses 381. Neither is a large number against 12,700, and if the misses were spread evenly through the day neither would matter to any decision you take.

They are not spread evenly, and this is where the headline figure stops being useful. On a trade show the arrival curve is brutally front loaded. Say 45 per cent of the day's arrivals come through in the 90 minutes from 09:30 to 11:00, which is 5,715 people, leaving 6,985 across the rest of the day.

Suppose the sensor runs at 99.3 per cent in the quiet hours, when people arrive in ones and twos with clear space around them. That accounts for 6,985 times 0.007, which is 49 missed. The day total of 381 missed then leaves 332 misses inside the peak window, and 332 out of 5,715 is an accuracy of 94.2 per cent for that window.

One instrument, one day, and two very different numbers: 99.3 per cent when nothing depends on it, 94.2 per cent during the ninety minutes the safety officer and the entrance manager are both watching.

The same 332 people read differently depending on which report they land in. Against the day's 12,700 they are 2.6 per cent, which nobody would query in an attendance figure. Against the peak window's 5,715 they are 5.8 per cent, and against the single busiest 15 minute bucket they can be most of the discrepancy in it. One number, three contexts, and the only one that gets published is the flattering one.

There is a consolation in that, provided you say it out loud. A bias of this kind is stable across editions, because the arrival curve and the door geometry barely change. So a year on year comparison of door totals holds up even while the absolute level runs low, which makes the sensor a decent instrument for direction and a poor one for a headline attendance claim.

Why do the misses cluster at the busiest minute?

Because the failure mode and the peak have the same cause.

An overhead height profile separates two people when there is a gap between them. Crowd density removes the gap. At the opening surge, people arrive in coach parties and conference groups, walking shoulder to shoulder through an aperture sized for two, and the returns merge into a cluster with one peak instead of two. Trolleys, wheeled cases and umbrellas add shapes the classifier has to make a decision about, and those also arrive disproportionately at the start of the day.

The same mechanism makes the error one-directional. Merging two people into one cluster undercounts. Splitting one person into two clusters is rarer, because a human body seen from above is a single connected mass. A crowded lidar count is a low count, systematically, and the shortfall grows with density.

That directional property is useful. If your running occupancy figure is built from these reads, the entry side is biased down at exactly the hour occupancy is climbing fastest, and the effect works against the upward drift that missed exit reads produce. Two errors in opposite directions do not cancel to zero in any reliable way, and unpicking them is its own job.

Which accuracy figure will your dashboard actually show?

Not the per-crossing one. Dashboards show buckets, usually 15 minutes or an hour, and bucket accuracy is a different measurement.

The 2020 study reports both, which is why it is worth reading properly. Alongside the better than 97 per cent disaggregate figure, the authors give average absolute percentage deviations at 15 minute intervals of 1.6 to 4.3 per cent across their two test locations. On a 15 minute bucket holding 900 arrivals, 4.3 per cent is 39 people.

The two figures answer different questions and neither one is derivable from the other. A sensor can get almost every individual crossing right and still show a wobbly quarter-hourly series if its timestamps drift or if it resolves a passage a few seconds late across a bucket boundary. It can also get a steady bucket series out of compensating errors.

So when a supplier quotes accuracy, ask which of the two it is, at what interval, and at what flow rate. If the answer is a single percentage with no interval attached, it is a per-crossing figure from a low flow test. Turning that into a number you trust means running your own comparison, and the two-observer design for that is the validation you can do in a morning in C23.

What does lidar buy that a camera cannot?

Mostly a different privacy position, and it is a real one.

Rukundo and colleagues surveyed sensing technologies for autonomous retail in an arXiv paper posted in March 2025, covering vision, RFID, weight sensing, vibration detection and lidar together. Their summary of lidar is that it builds a depth map of an entrance or an aisle, and that because it does not capture facial details it is privacy preserving while still being accurate enough to count customers. That is the honest case for it, and for an exhibition organiser it removes an entire class of conversation with a venue, a works council and an exhibitor whose stand happens to sit in frame.

The second thing it buys is indifference to light. A hall at 08:00 with the house lights half up, a hall under stand lighting, and a hall with sun through a glazed concourse are the same scene to a range sensor and three different scenes to a camera.

What it costs is context. A depth cluster has no attributes, so it cannot tell you a person from a contractor with a pallet truck beyond what the height profile implies, and it cannot recognise the same person at a second doorway. Density map counting from video carries a different set of failure modes and a different privacy analysis, which is the camera side of the same decision in C20.

Where this stops

Everything above is about a doorway, and a doorway is the friendly case. The background is static, the aperture is narrow, the mounting height is controllable, and people cross a line in a direction.

The middle of a hall has none of that. Stands change shape between builds, so there is no stable background to subtract. Mounting points are wherever the rigging plan allows. People mill rather than cross, so there is no counting line and no direction. Accuracy figures earned at a door transfer to open floor about as well as a corridor calibration transfers to a stadium, which is to say they do not, and any proposal that quotes a doorway figure for a floor deployment should be sent back with that question.

The second limit is that lidar counts crossings and nothing else. An anonymous crossing cannot be joined to a badge, a registration record or a session, so it can size a flow and it cannot answer a question about who. Building a hall's instrument list around the decisions rather than the hardware is the selection exercise in C24, and it is the first thing to settle in any attendee analytics plan.

The first step costs one afternoon of last edition's data. Pull your door counts in 15 minute buckets, find the bucket with the highest arrivals, and compute what share of the day arrived in the busiest 90 minutes. If that share is above 40 per cent, your sensor's quoted accuracy is being earned in the wrong hours, and the number you should be asking the supplier for is the one measured at your peak flow rate.