3D sensor, camera, infrared and lidar or Wi-Fi: which technology counts people how accurately?

Every evaluation starts with the same question: which technology should go on the ceiling? For people counting that holds up in crowds, 3D stereo vision is, in our assessment, the most accurate of the common methods. Infrared is enough for daily totals at a quiet door; Wi-Fi and Bluetooth count devices rather than people. Yet the question is usually answered with a percentage the vendor supplies, and that answer is usually wrong.

This article works through four widely used technologies. For each it gives the operating data an IT department will ask for anyway, and at one point it does a calculation almost nobody does: what a small counting error does to your occupancy figure.

You choose the technology once. You live with the consequences for ten years.

Why accuracy claims are not comparable

Practically every vendor quotes accuracy above 98 or 99 per cent. Those numbers sit side by side as if something could be inferred from them. Nothing can, because they were produced under different conditions.

Accuracy in people counting is always stated per crossing and always under defined conditions. Measuring 98 per cent while individuals walk through a wide door at a calm pace, and measuring 98 per cent while a school class pushes through that same door at once, are two entirely different measurements.

The questions that make an accuracy figure usable:

• At what traffic density? People per minute per metre of door width. The error does not rise linearly with density but disproportionately, as soon as people start occluding one another.
• How was the reference established? Manual counting by two independent people with reconciliation, or video re-counting? A reference that carries its own error cannot prove 99 per cent.
• Over what period? A thirty-minute measurement at midday says nothing about closing time.
• Broken out by entry and exit? This is the most important question, and it is almost never asked. The next section explains why.
• Were groups, prams, wheelchairs and shopping trolleys counted? And if so, as what?

Without a test protocol, a percentage is marketing. With one, it is a commitment you can verify at acceptance.

The calculation almost nobody does

If you only count at doors, you do not know occupancy directly. It is computed: all entries minus all exits. That subtraction hides a problem that appears in no brochure.

A uniform deviation is harmless. If a system counts both entries and exits two per cent low, the difference is two per cent low as well. With 200 people present that is 196. You can work with that.

The trouble starts as soon as the deviation differs between entrance and exit. And that is the normal case, not the exception: people arrive in groups and leave individually. The entrance is wider than the exit. Morning light differs from evening light. There is a queue at the exit and none at the entrance - and what a queue does to the numbers is something we have taken apart in detail elsewhere.

An example with realistic numbers. A floor records 5,000 entries and 4,800 exits by 6 pm. Actually present: 200 people.

A single percentage point of difference between entrance and exit turns a two per cent deviation into a twenty-six per cent one. The reason is elementary: the error grows with total traffic, while occupancy is a small difference between two large numbers.

And it gets worse the larger the floor. At 20,000 entries and 19,500 exits, meaning 500 people present, the same two-versus-one per cent constellation already produces a 41 per cent deviation.

Occupancy is a small difference between two large numbers. Small errors in the large numbers are large errors in the result.

Two practical consequences follow.

First: an accuracy figure not broken out by entry and exit says nothing about occupancy measurement. Ask for the breakdown.

Second: if you genuinely need occupancy, dwell time or utilisation, ask when the balance starts again. Every error that enters the difference stays in the figure until the next zero point. A fixed daily zero point at night, when the floor is empty, keeps yesterday's deviation out of today's value. And if you also cover zones on the floor, you see not only how many people are there, but where they spend their time and for how long.

Technology 1: 3D stereo vision

Two lenses at a fixed distance deliver two slightly offset images of the same scene. From the shift between them, a processing unit inside the device calculates the distance to every pixel. The same principle by which two eyes see in three dimensions.

The gain is the third dimension. The sensor does not see a surface with blobs on it, but a height map. Two heads side by side are two maxima and therefore two people. An adult and a child differ in height. A shopping trolley has no head profile. Exactly the cases where simpler methods fail are resolved by the depth information itself.

All image processing happens inside the device. What leaves it is text: counting events, positions and attributes derived from them such as body height, each with a timestamp. No image, no video stream.

Operating data your IT department will ask for (manufacturer specifications)

• Power: Power over Ethernet to IEEE 802.3af, class 0, nominal 48 V. Maximum consumption 7.5 to 13 W depending on model. No separate mains supply at the ceiling, no socket, no electrical work at the mounting point.
• Cabling: one cable per sensor, shielded Cat-5e or better, Cat-6 for some models, up to 100 m, Gigabit Ethernet over RJ45.
• Network: IPv4 and IPv6, DHCP, HTTPS and, usually decisive for approval on a corporate network, IEEE 802.1x. Without port authentication a device will not be admitted to the network in many organisations.
• Data delivery: REST interface with Swagger documentation, plus active data push over HTTP(S), FTP(S), SFTP, MQTT(S), TCP or UDP. MQTT is the reason such sensors attach to a building management system without bespoke development.
• On-device storage: up to three years of counting data, depending on the number of configured counters. A network outage therefore does not create a gap in the record.
• Light: from 2 lux. That is less than dimmed museum lighting. A method that needs brightness is ruled out in an exhibition.
• Mounting height: depending on model 2 to 6 m for ordinary rooms, 6 to 20 m for halls, atria and station buildings. The height determines the field of view and therefore how many devices a floor needs.
• Mounting tolerance: roughly 15 degrees of tilt in one axis, 5 degrees in the other. A ceiling need not be perfectly level, but it cannot be arbitrarily skewed either.
• Environment: indoor models for 0 to 45 °C, 20 to 80 per cent relative humidity non-condensing, ingress protection IP40. For outdoor use there are dedicated models protected against water and dust, for minus 33 to plus 40 °C.
• Reliability: a calculated MTBF of over 30 years to IEC 61709 (manufacturer specification), depending on model and ambient temperature. The warmer the surroundings, the higher the calculated failure rate.

What that means for planning: large floors need correspondingly many devices, and the mounting points have to be planned. This is not a device you hang somewhere and hope for the best. Covering a floor without gaps requires overlapping fields of view, and those follow from ceiling height and layout.

Technology 2: camera with facial recognition

A video camera records, software evaluates. Video analytics comes in many forms, from simple counting in the image to facial recognition. Facial recognition is the case considered here because it goes furthest: with it, people can be uniquely identified and recognised again, which in theory also yields dwell time and repeat visits.

Three points argue against it.

The perspective. Cameras usually hang at an angle, not vertically. Anyone standing behind someone else is occluded. Precisely at high density, where the data would be worth most, the gaps appear.

The infrastructure. Video has to cross the network and land on storage. With several cameras over months this becomes an installation of its own, with its own maintenance and its own failure risk. The data load is orders of magnitude above a device that sends only text.

The legal framework. Facial features are biometric data. Where they serve to identify someone uniquely, they fall under Article 9 of the GDPR as a special category of personal data; in Switzerland they are sensitive personal data under the revised FADP. Using them generally means a data protection impact assessment, a deletion concept and notice to the people affected, and in the EU the question of a legal basis on top. That is not a formality, it is a second project alongside the counting project.

Storing faces solves a counting problem by purchasing a data protection problem.

Technology 3: light and laser, meaning infrared and lidar

Both work with light, but in different ways. An infrared light barrier detects that a beam has been interrupted. Lidar emits laser pulses and measures the time until they return. That yields a distance, and many distances yield a shape.

Infrared is the classic in the door frame: often already installed and proven for decades. What it lacks is accuracy. If the system works with single beams, a larger person occludes a smaller one completely, and two side by side become one. For a daily total at a quiet door that is fine. For an occupancy calculation, per the maths above, it is not.

Lidar uses a focused laser, reaches further and measures considerably more precisely. A high-quality sensor covers a substantial area.

Both share a sensitivity to their surroundings: fogging and condensation on the optics, and rain, snow or fog outdoors, disturb the measurement. Additional attributes such as view direction generally cannot be retrofitted.

A genuine advantage: nobody can be uniquely identified from an interrupted beam or a time-of-flight reading. In data protection terms these methods are therefore uncritical.

Technology 4: Wi-Fi and Bluetooth

Sensors pick up the radio signals of devices people carry and try to recognise them again by MAC address. The method has been overtaken from several directions at once.

The sample is wrong. Anyone carrying two phones, work and private, is counted twice. Children and many older people carry none and are missing entirely. What gets counted is devices, not people, and the ratio between the two is neither known nor stable.

The basis was deliberately removed. Operating systems now assign randomised MAC addresses, per network and when scanning for networks, in some cases rotating them regularly. That is a necessary privacy measure by the manufacturers, and it renders re-identification across networks and days useless. An installation built on it loses value with every system update, without anyone having done anything wrong.

And in data protection terms it is delicate. A MAC address is a unique identifier, and datasets exist that link such identifiers to individuals.

The four side by side

The table is a qualitative assessment, not a measurement. It does not replace the test protocol this article asks for.

Data protection is an architectural decision, not an afterthought

Switzerland has had a revised Federal Act on Data Protection since 1 September 2023; the EU has the GDPR. The two take different approaches: the GDPR requires a legal basis for any processing of personal data, while the revised FADP requires a justification once processing infringes someone's personality. They agree that biometric data which uniquely identify a person are particularly sensitive.

For a tender this creates an uncomfortable asymmetry. If a method can identify people, you have to keep explaining why you use it, how you protect the data, how long you keep it and how you inform those affected. If a method cannot identify anyone in the first place, that discussion becomes much shorter: in the project, with the employee representatives, in audit, and towards your visitors.

The most robust data protection concept is the one where identifying data never comes into existence.

That is why the 3D sensors we use work without facial recognition. The evaluation happens inside the device, and the image stays there. What leaves is text: counting events, positions and derived attributes, linked to a session-bound identifier rather than an identity. In counting operation no image leaves the sensor, and ACS stores no images.

Checklist for your tender

Seven questions that separate a commitment from a percentage:

1. Accuracy broken out by entry and exit, with test protocol and stated traffic density.
2. Is only the door line covered, or also zones on the floor, and when does the occupancy balance start again? That determines how far occupancy can drift and whether dwell time can be measured.
3. What data leaves the device? Image, video, or text only?
4. Does the device support IEEE 802.1x? Without it, approval on the corporate network fails.
5. How many devices does your layout need at your ceiling height? That answer belongs in the planning, not in the rollout.
6. How long does the device store data itself if the network drops?
7. What interface is there, and is it documented? A REST interface with Swagger is an integration project of days; an undocumented one is a project of weeks.

How we do it

The ANALYSIT Counting System uses 3D sensors at the entrances and, where you need them, zones on the floor. It reports by the minute how many visitors arrive, looking back over 730 days, plus live occupancy, for each zone too, and how long they stay where. The running occupancy balance starts again every day at 3 am local time.

No image, no facial feature and no device identifier leaves the sensor. ACS receives no identifying data: the per-minute counter carries no personal reference, and where paths and dwell time are concerned, ACS works with a session-bound, pseudonymous identifier from the sensor rather than an identity.

If you want to know how many sensors your floor needs and what can be measured on it, talk to us. In the site survey, layout, ceiling heights and door widths are recorded. From them follow sensor model, mounting height and number of devices, and the counting lines and zones are set on the floor plan.