People counting technologies compared: beam, thermal, 3D stereo, time-of-flight, AI video and Wi-Fi (2026)

Accuracy figures are the ranges vendors publish for their own products under good conditions; the two sensors we sell are marked. Mounting heights are typical published ranges. “Privacy” describes what the sensor can capture by design, which is what matters legally, not what a particular installation is configured to keep.

The original people counter, in use since the 1980s and still the most-installed type. An infrared transmitter on one side of the doorway shines at a receiver on the other; when a body interrupts the beam, the receiver registers a count. That is the whole mechanism, and its simplicity is both its strength and its limit.

Variants

Reflector beams use one active unit and a mirror; cheaper, but the reflector is bulky, easily knocked and prone to sunlight errors. Transmitter/receiver (Rx/Tx) beams are smaller and more reliable; every beam we have sold uses them. Directional beams put two beams a few centimetres apart on each side: A then B is an entry, B then A an exit. Wireless beams pulse the beam to save power and run on batteries for most of a year; they are the entry point for most first-time buyers, at the price of a battery routine and a radio link that is very stable in 2026 but never quite as certain as a cable. For a chain of a hundred locations, that maintenance load is the real argument for spending more.

Our beam: the PEARL — an Rx/Tx pair on AA batteries over 2.4 GHz Wi-Fi, peel-and-stick, in/out, upwards of 98% on adult traffic, $499.95 with software from $0. It is the right sensor for a standard door and the wrong one for a wide, busy or glass entrance, which is why we also make the next thing.

The same sensor that switches on a security light. Mounted over the lease line, its detection zone is narrowed to a virtual line a few inches deep; a warm body moving through it triggers a count after a reset time of half a second to two seconds. It is inexpensive and has no image, and in a climate-controlled mall corridor with free-flowing traffic it behaves like a ceiling-mounted beam.

Its failure list is the beam’s plus three: it is upset by rapid temperature change, so an entrance to the outdoors on a cold day generates phantom counts; it cannot filter children or carts because it sees heat, not height; and a person who dawdles or stands in the zone can be counted repeatedly as the sensor resets. Battery-powered PIR units add ceiling-height battery changes that knock the unit out of alignment. We stopped recommending PIR for most retail doors a decade ago; it survives in some low-cost occupancy sensors.

A low-resolution thermal camera looks straight down and sees people as warm blobs moving across a cooler floor. Firmware tracks each blob and counts it when it crosses the line. Pioneered in the UK in the 2000s, thermal is today the technology behind the battery-and-cellular counters on the market, several of which pair a thermal array with machine learning.

Verdict: a correctly calibrated thermal counter is a system you can trust, and for a dark entrance, an outdoor gate or a site with no network and no cable it can be the best choice. For a bright, busy retail door with families and carts, stereo does more.

A single camera looks down; software separates moving foreground from a learned background and counts the blobs that cross a line. It was the third generation of counting (roughly 2012 onward) and it is now largely superseded, for one reason: a single lens has no idea how tall anything is. A shadow moving across the floor, a shopping cart, a child and an adult are all just moving pixels. Accuracy is fine in a quiet, evenly lit corridor and collapses in crowds, glare and hard shadows.

The same applies, with extra problems, to re-using security cameras. CCTV is mounted at an angle to see faces, which is the wrong geometry for separating people who overlap; it records identifiable video, which brings the full CCTV rulebook on signage, retention and access; and the analytics integration usually costs more than a dedicated sensor. Modern AI helps 2D video considerably (see below), but where a depth sensor is affordable, it wins.

Two lenses a few centimetres apart see the scene from slightly different angles; from the difference between the two images the sensor computes the height of every point below it, producing a depth map in which people are head-shaped objects at 1.5–1.9 m and carts are flat objects at 1 m. Each object is tracked across the field of view and counted once, in the direction it crossed the line. “3D” and “stereo” mean the same thing in this industry; the term was a marketing choice. Since about 2010 stereo has been the technology to beat for retail, and nearly every enterprise vendor, ourselves included, builds on it.

Ours: the 3D Scope II LC — stereo with on-device AI and HDR, 99% rated with a 95% floor guaranteed in writing for two years, one PoE cable under 8 W, coverage up to 8 × 8 m per sensor (8 × 22.8 m with two), in/out, adults and children, carts, groups, wheelchairs, U-turns, staff exclusion, video audit on request; $1,149.95, outdoor model $1,650. Where we lose: ceilings above 6 m, and anyone who needs age and gender, which we deliberately do not estimate.

A time-of-flight sensor emits pulses of infrared light and times their return from every point below; the round-trip time gives distance, and distance from the ceiling gives height. The result is a depth map very like a stereo sensor’s — heads, carts and children distinguished by height, groups separated, direction known — but produced without any colour image at all. Popularised by the Microsoft Kinect, ToF was “new and promising” when we last wrote about it in 2016. It has since matured into a real alternative to stereo from several specialist vendors; LiDAR units apply the same principle with scanning lasers over larger areas.

“AI” on a 2026 datasheet means the sensor runs a neural network on the device to classify what it sees, rather than applying hand-written rules. On a stereo or ToF sensor the network works on depth data: it learns what a head, a cart, a wheelchair, a stroller and a child look like from above, so the sensor can count carts separately, count a wheelchair user as a person, tell a tall child from a short adult better than a fixed height threshold can, and separate a tight group. On thermal sensors, machine learning has closed much of the accuracy gap to cameras. On 2D video, AI object detection rescued a technology that background subtraction had left behind.

Three things it does not mean. It does not mean video is being sent to the cloud — the whole point of edge processing is that only counts leave the device. It does not mean face recognition; a counter trained on head shapes from above never sees a face. And it does not mean accuracy is solved: an AI sensor mounted over a door swing, in the dark, or above its rated height still counts badly. The useful questions are the same as ever — what does the sensor capture, what leaves it, and what is the guaranteed floor.

Where AI adds capabilities you should think twice about: age and gender estimation requires real images of faces, is biometric processing under most privacy regimes, and is a different product with a different legal footing. Some vendors publish accuracy figures for it; we do not offer it by choice.

Between about 2012 and 2015 a sensor that listened for the Wi-Fi probe requests every smartphone broadcasts, and counted the unique hardware addresses, looked like a way to count people, measure dwell time and recognise repeat visitors at once. Its problems were always there — it counts devices, not people (none for a child, two for a rep with two phones), its positioning is poor to a few metres so it counts the pavement as the store, it needs Wi-Fi switched on, and it collects a device identifier that regulators in the EU and Quebec treat as personal data. Then the phone makers ended it.

Apple (from 2014) and Android 6 (2015) began randomising the address in probe requests; iOS 14 (2020) introduced a private Wi-Fi address per network, and iOS 18 (2024) added an option that rotates the private address every two weeks and turns it on by default for open networks (Apple). Unique-device counting became unique-random-string counting. Vendors that survived now sell Wi-Fi analytics as statistically corrected estimates or as opt-in guest Wi-Fi with consent. That can be useful for dwell time and zone flow in a mall. It is not a door count, and no serious vendor sells it as one any more.

The radio technology that is useful at a door in 2026 is ultra-wideband, in the other direction: staff carry a UWB tag and a receiver on the sensor tells the counter to ignore them (how our staff exclusion works). Consent is built in, because the only devices tracked are the ones you handed out, and the tags identify no individual.

Every overhead sensor gives direction for free, so the question only really arises for beams, where a second beam adds cost. Our answer has not changed in twenty years: for traffic analysis, no; for occupancy, yes, and then only with the best sensor you can afford.

Over any period, the people who came in and the people who went out are the same people. Staffing a half-hour on IN counts, or on total beam counts divided by two, produces nearly the same rota; staffing on “people currently in store” is less reliable, because it depends on when within the half-hour they left (the worked table). Conversion rate uses IN counts and is indifferent to OUT.

Occupancy is the exception and the trap. It is cumulative IN minus cumulative OUT, so every counting error persists. A 95%-accurate sensor at a door seeing 100 people an hour makes about five errors an hour; if they lean one way, and they do, the system believes there are 25 phantom people in the store by closing and 60 by Sunday night. That is arithmetic, not a fault. Live occupancy is a legitimate use — for capacity limits, gyms, study halls, event floors — but it needs a high-accuracy sensor on every entrance, a daily reset, and a manual correction button, which is how our Real-Time plan is built.

The “group counting” feature

Some sensors offer to count people who enter within a set distance of each other as one “shopping unit”, which raises the reported conversion rate. The distance rule cannot know whether two people are a couple, colleagues or strangers who arrived together, so it is arbitrary at best and flattering by design. We have offered it and we advise against it: count people, know that some of them shop in groups, and measure your improvement against a consistent baseline.

Our own specifications come from our product pages and help centre, linked in place. Industry accuracy ranges are the figures vendors publish for their own products, summarised without naming them; no competitor is quoted or linked on this page. Named comparisons live on the comparison page.

How this guide was made

This page replaces the three-part “How customer counter technology works” series we published in 2016, keeping what has stayed true — the 54-inch rule, the beam’s failure list, the thermal standing-still problem, the occupancy drift arithmetic, the scepticism about group counting and Wi-Fi — and updating what changed: time-of-flight matured, edge AI arrived, Wi-Fi counting ended, and prices fell. SMS Storetraffic sells two of the technologies described; we say so where they appear and name other vendors where they fit better. Corrections via our contact page.