How Does a People Counter Work? Sensor or Camera?

How does a people counting device work? Differences between infrared, thermal, 3D stereo, ToF, radar, Wi-Fi and video analytics, installation conditions, KVKK and cost factors.

Translated from the Turkish original · Türkçe aslı

A people counter watches a defined line or area at the entrance and counts every person who passes, together with their direction. It does this by detecting a broken beam (infrared beam), body heat (thermal), by measuring depth (3D stereo, ToF), with radio waves (mmWave radar), from phone signals (Wi-Fi/BLE), or by processing camera footage with AI. The choice is driven by the physics of the entrance and by what data you need beyond the count.

What is the real difference between people counting technologies?

Every people counting system does the same four jobs: it detects the person, determines in which direction they cross the counting line, filters out what should not be counted (staff, people waiting at the door, shopping trolleys) and turns the result into an hourly report. The technologies diverge at the first step; the detection method determines how consistent the count stays with group entries, changing light and wide doorways. Even so, the accuracy gap between well-installed systems in the field is usually smaller than people think; the real difference is how much data you get from the same investment. A beam counter gives you only the number of entries. A system that processes images can also calculate queue length, in-store routes and conversion from the same stream. A count alone does not produce a decision; knowing how many people came in does not explain how many left without buying, or why.

How do infrared beam and thermal counters work?

An infrared beam counter sets up an invisible beam between a transmitter and a receiver (or a transmitter and a reflector) mounted on either side of the door; each time the beam is broken, the counter goes up by one. For direction, two parallel beams are used and the system checks which one is broken first. It is cheap, battery-powered models exist and installation is simple. Its weakness comes from its physics: two people entering side by side break the beam at the same time and are counted as one; a customer standing in the doorway or a pushchair waiting in front of the beam disrupts the count. The wider the entrance and the longer the beam, the bigger the problem. It cannot tell staff apart.

A thermal counter looks down from the ceiling and uses a low-resolution sensor array that detects body heat. It is independent of light, works in the dark and, because it does not produce an identifiable image, it is straightforward from a privacy standpoint. When the ambient temperature approaches body temperature (at entrances with open doors in summer, or under heaters and air curtains), the contrast between person and background shrinks and detection becomes harder. Its ability to separate people walking very close together is limited by the low resolution.

What do 3D stereo and ToF sensors do differently?

A 3D stereo sensor looks at the same scene through two lenses set a distance apart, like human eyes, and calculates the distance of every point from the shift between the two images. Looking down from the ceiling, the sensor thus builds a depth map: every bump rising from the floor is a person, and their height is measured. Thanks to a height threshold, shopping trolleys and small children can be separated out, people walking side by side appear as separate bumps, and shadows and changes in light barely affect the count. The price is coverage: the area the sensor sees is limited by the ceiling height, wide entrances need several sensors combined, and a dedicated piece of hardware that serves only the counting purpose has to be mounted.

A ToF (time of flight) sensor measures depth another way: it emits modulated infrared light and calculates distance from the time the light takes to return, in practice from the phase difference. Because it generates its own light, it works in the dark and does not need to match two images. Entrances in direct sunlight and very bright, reflective floors can disrupt the measurement. Both technologies require overhead mounting and do not see faces; when mounted correctly, they are among the sensor families that give the most consistent counts with group entries.

How reliable is counting with mmWave radar and Wi-Fi/BLE?

Millimetre-wave radar sends out a radio signal and derives the distance, speed and direction of targets from the reflection. It produces no image and is unaffected by darkness, smoke and fog. These properties make it attractive for measuring occupancy in areas where a camera is not appropriate, such as toilets, changing rooms or patient rooms. It can struggle to separate people standing very close together, and reflections from metal surfaces can create ghost targets. Because there is no image, it is not possible to go back and verify a counting error or to keep an evidential record of the event.

Wi-Fi and BLE based systems actually count phones: they listen to the Wi-Fi probe requests and Bluetooth advertisements that devices emit. A visitor with no phone, with Wi-Fi turned off, or carrying two devices skews the result; the MAC address randomisation of modern phones causes the same device to be counted more than once or a repeat visit to go unrecognised. Because the signal passes through walls and glass, people walking past the shop can also be detected. This method is better suited to estimating density and trends between zones than to exact entry counts.

How does people counting with video analytics work?

In camera-based counting, a deep learning model finds the people in each frame, a tracking algorithm gives each person a short-lived track ID, and when the person crosses the defined line they are counted along with their direction. The best results come from an overhead camera above the door; if an existing security camera has a suitable angle, it can be used as well. The counting accuracy verified in the field for the people counting module of EscherSense is 96.2%. The difference of working with images begins after the count. The same infrastructure derives heatmap and route analysis from the cameras inside the store, measures the checkout queue, calculates conversion by matching counts with the number of POS receipts (conversion measurement accuracy 92.8%), and, from a separate camera point, anonymously measures the age and gender distribution of the visitor audience (demographics measurement accuracy 96.2%). A questionable count can be checked against the footage of that moment. Its weaknesses are changing light, backlight and people occluding each other in crowds; an overhead angle reduces most of these problems. We describe the camera counting flow and the conversion calculation in detail in the people counting system article.

The table below puts the detection principle of each technology and its strengths and weaknesses in the field side by side.

How do ceiling height, entrance width and light affect installation?

Every overhead sensor (3D stereo, ToF, thermal and overhead camera) has a mounting height range specified by the manufacturer. If the ceiling is below this range, the floor area the sensor sees narrows and a wide door cannot be covered with a single device; if it is above the range, people appear smaller and depth and heat separation weaken. Entrance width is the main factor that determines the number of devices: a narrow shop door is counted with a single device, a multi-leaf shopping mall entrance with several linked devices, and the same person must not be counted twice where the devices' fields overlap. Light affects optical systems. At glass entrances that get sunlight, shadows and backlight that change during the day challenge the camera, and direct sunlight challenges ToF; the thermal sensor is unaffected by shadows and backlight, but a floor warmed by the sun can reduce the temperature difference between person and background; radar is unaffected by light. The mat in front of the door, revolving doors, the movement of automatic doors and promotional stands set up at the entrance are also practical details that determine where the counting line goes.

How are group entries and staff separated out?

Group entries are the biggest source of counting error. A single-beam counter counts two people side by side as one; depth-measuring sensors and overhead cameras see people individually and so handle this far better. A child being carried or a couple walking arm in arm is challenging for every technology. Retail analysis raises a further question: is a family of five five visitors, or a single shopping group? Using the number of groups in the conversion calculation can be more meaningful; some systems flag people moving together as a group.

Separating out staff is critical for the conversion rate, because when staff crossings are added to the denominator, conversion looks lower than it is. Sensor-based systems usually do this with a badge or tag worn by staff; staff who forget their badge are counted as visitors. Camera-based systems can use visual and behavioural cues such as uniforms, exits from staff areas or moving behind the checkout. Whichever method is chosen, accuracy should be measured with the staff filter switched on, by comparison with manual counts taken during busy hours.

What does each technology mean under KVKK?

Infrared beams and low-resolution thermal sensors do not produce an identifiable image, so the personal data risk is low. Most 3D and ToF sensors process only depth data, but some models can also record colour images; which data is stored where should be asked device by device. Wi-Fi and BLE systems work with device identifiers, and these identifiers can count as personal data to the extent that they can be linked to a person. Camera footage is personal data, and the design is governed by the principle of proportionality: processing footage by default on a device inside the store without sending raw video to the cloud, not using face recognition, keeping the output as hourly counts, and making sure the privacy notice clearly states the counting purpose together strike this balance. We cover the general framework of KVKK (Türkiye's Personal Data Protection Law) in the KVKK and image processing article; this information is not a substitute for legal advice.

What does the cost of a people counting system depend on?

People counter prices cannot be compared on a single figure, because the total cost is determined more by the installation conditions than by the device itself. The main items are:

• Number and width of entrances: each entrance needs at least one device, wide entrances need several.
• Infrastructure: cabling to the ceiling, PoE switches, mounting with a lift platform if needed, and the in-store network.
• Software: the dashboard and reporting often come with an annual licence per device or per store.
• Integration: pulling hourly receipt data from the POS and feeding ERP and BI tools.
• Validation and maintenance: calibration against manual counts after installation, and readjustment when the store layout changes.

Video analytics that uses existing cameras reduces the new sensor and cabling items, but adds processing hardware. The right yardstick is not the price per device but the cost per metric obtained; we go through the items one by one in the AI camera system cost article.

Which should you choose for a store, shopping mall, factory or event?

In a single-entrance store the goal is usually the conversion rate; you need a system that can separate group entries and filter out staff. If the store already has cameras and you also want to understand queues and in-front-of-shelf behaviour alongside the count, video analytics yields more data from the same investment. If the number of entries alone is enough and the budget is tight, an infrared beam is a starting point; you should know from the outset that the count will fall short with group entries.

A shopping mall has many wide entrances, car park connections and floor transitions. Consistency matters here: if the same technology and the same counting rule are not used at every entrance, floor and corridor comparisons show measurement differences rather than store performance. Overhead depth sensors or overhead cameras suit wide entrances, and zone-based counting with existing cameras suits floor and corridor distribution.

In a factory, the turnstile already keeps a per-person entry–exit record through badge scanning; an extra sensor for counting is often not needed. Here the camera answers a different question: whether two people passed on one badge, whether anyone went around the turnstile, and how many people remain in an area during an emergency. At events the installation is temporary and the real need is real-time occupancy; with wireless sensors or temporary cameras, entries and exits are counted separately, the number of people inside is continuously calculated, and an alert is generated when a threshold is exceeded. At outdoor events, changing daylight and wide, undefined entrances are the biggest challenges.

CX Teknoloji's people counting solution works with existing IP cameras, can process footage inside the store and does not use face recognition; the same infrastructure also carries the heatmap, queue and demographics modules. In a single-store pilot, the actual crossings during specific hours are counted manually and compared with the system's count, so accuracy is measured at your own door. To have your existing camera angles assessed or to arrange a demo, get in touch with us.

Frequently asked questions

What is the difference between a people counting sensor and a people counting camera?

Sensors (infrared, thermal, 3D stereo, ToF, radar) are single-purpose devices designed only for counting; most do not produce images and are straightforward from a privacy standpoint. Camera-based counting processes images with AI; it can also derive queue, route, conversion and demographics from the same stream, and a questionable count can be verified from the footage. The accuracy gap between well-installed systems is usually smaller than assumed; the deciding factor is the scope of data you need.

What determines people counter prices?

The price is determined more by the installation conditions than by the device. The number and width of entrances determine how many devices are needed, ceiling height and cabling determine the mounting labour, the dashboard and reporting licence determines the annual cost, and POS and ERP integration determines the project scope. In systems that use existing cameras, the new sensor and cabling items shrink and processing hardware is added. When comparing quotes, look at the total cost of ownership and the number of metrics obtained.

Can an existing security camera be used for people counting?

If the camera sees the entrance from above or at a steep angle, it can often be used. With cameras looking at the door from the side, accuracy drops at busy times because people occlude each other; in that case adding an overhead camera above the door gives better results. ONVIF-compliant IP cameras that provide an RTSP stream are generally sufficient on the software side; what really matters is angle, resolution and light.

Does a people counting system count children and shopping trolleys?

It depends on the technology. An infrared beam counts anything that breaks the beam; shopping trolleys and pushchairs can end up in the count. 3D stereo and ToF sensors measure height, so trolleys and children below a certain height can be separated out with a threshold. A camera-based system recognises people as a separate class and so does not count trolleys; whether children are included in the count is defined as a rule according to the business's preference.

Does people counting with Wi-Fi give accurate results?

It is not considered a suitable method for exact entry counts. Wi-Fi and BLE systems detect the signals phones emit; they miss visitors without a phone or with Wi-Fi turned off, may count someone carrying two devices twice, may see the same device several times because of MAC address randomisation, and may also detect people walking past the window. They can be used for density trends between zones and rough dwell time estimates.

Is people counting within the scope of KVKK?

It depends on the technology used. With beam and thermal sensors that do not produce identifiable images, the personal data risk is low; camera footage and device identifiers that can be linked to a person, however, can count as personal data. For camera-based counting, it is recommended not to use face recognition, to process footage on site, to keep the output as anonymous counts and to update the privacy notice. This information is general in nature and is not a substitute for legal advice.

How is people counting accuracy tested?

The actual crossings in specific time slots are counted manually from the video recording and compared with the system's count for the same slot. The test should include busy and quiet hours, group entries and staff crossings, and be carried out with the staff filter switched on; entries and exits are measured separately. The value measured at your own door, rather than the manufacturer's lab figure, should be the benchmark, and the test should be repeated when the store layout changes.