People counting explained: how beam, thermal, 3D and AI video counters compare, which number you need, where to mount the camera and how to test accuracy.
People counting answers three different questions: how many passes, how many unique visitors, and how many people are inside right now. Know which one you need first.
Every method has a blind spot. Beams miss people walking side by side, Wi-Fi misses people without phones, and cameras need the right angle and light.
The CCTV cameras you already own can often count people with AI video analytics, with no new sensor on the ceiling.
Placement decides accuracy more than the brand of counter.
Never trust a quoted accuracy figure. Test any counter against a hand count at your own busiest door.
Counting people does not require knowing who they are. Recognising a returning visitor is more sensitive, so choose it deliberately.
Counts pay off when they trigger action, such as an alert at capacity.
Search for people counting and most results come from companies that sell one kind of sensor. Each explains why its own method is best. Few explain the three different numbers a counter can produce, which one your question needs, or how to check the figures yourself.
Fewer mention that the cameras already on your ceiling may do the job. This guide compares the methods honestly, then covers placement, testing and privacy.
People counting is the automatic measurement of how many people pass a point, visit a space or occupy it at a given moment, using sensors or cameras instead of a person with a clicker. A people counting system turns those counts into footfall, unique visitor and live occupancy figures for operations, safety and sales analysis.
A device that does this is called a people counter. Retailers use counts for conversion, facilities teams to plan space, and venues to stay within safe capacity.
People counters detect a person at a fixed point, usually a doorway, and record the crossing with a direction and a time. Each of the five common methods has a blind spot.
Infrared beam counters. A small sensor on one side of the door sends a beam across to the other, and each break counts as a person. They are cheap and simple. They cannot tell two people walking side by side from one, a single beam cannot tell direction, and a trolley counts too. Fine for a rough trend; poor for a figure you need to trust.
Thermal counters. A sensor on the ceiling looks straight down and detects body heat. It works in the dark and never sees a face, which makes it naturally private. It struggles near heaters, in strong sun or when the air is close to body temperature, and its low resolution limits how well it separates a dense group.
3D stereo and time of flight counters. An overhead sensor measures depth, either with two lenses like a pair of eyes or by timing reflected infrared light. Because it sees height, it separates adults from trolleys and splits tight groups well, which makes it a dependable way to count a single doorway. It needs a suitable ceiling, cabling and one unit per entrance, and it counts only its own small patch of floor.
Wi-Fi and Bluetooth counters. A receiver listens for signals from phones and counts the devices nearby. It covers a wide area, but it counts phones, not people: some visitors carry two, some carry none. Phones also now hide behind changing addresses. Apple says its private Wi-Fi address exists so that observers cannot easily relate one address to a device over time, and Android has used a randomised address by default since Android 10. That is good for privacy and bad for counting individual visitors.
AI video analytics. Software watches the feed from an ordinary IP camera, detects each person, follows them through the frame and counts them as they cross a virtual line or enter a zone. One camera can count a doorway, measure a queue, watch crowd levels and keep recording for security, often on cameras you already own. It needs a clear view and enough light, and because it processes images of people, it brings data protection duties that a thermal sensor avoids.
Choose a dedicated sensor when you need one doorway counted and nothing else. Choose a people counting camera with AI video analytics when you want counts, security, queues and crowd alerts from the same hardware.
A dedicated 3D sensor is hard to beat at a single, well defined entrance. It does one job well, but each new entrance means another device and another cable run.
A camera based approach wins with many entrances, open concourses with no doorway to mount over, or existing CCTV already watching the places you want to count. The same camera then serves security and operations.
Several camera makers, Hikvision among them, sell cameras with counting built into the camera itself. That suits a single shop. Across a larger estate, counts end up scattered across devices and brands, so many organisations count in the video platform instead.
Most disappointment with people counting comes from asking one number to answer another number's question.
Passes, or crossings. Every time anyone crosses the counting line, in or out, it counts. Passes answer "how busy was this entrance?" They overstate visitors, because staff and anyone who steps out to take a call are counted every time.
Unique visitors. Each person counts once per visit, however many times they pass. This is the figure that belongs in a conversion rate.
Live occupancy. Entries minus exits, right now. Occupancy answers "how many people are inside?" It matters for safety limits and staffing.
Footfall can mean passes or visitors, depending on who is reporting. Agree a definition before comparing two sites.
A simple test helps. If the question contains "right now", you need occupancy. If it asks how many people came, you need unique visitors. If it is about a door, passes will do.
To count unique visitors, the system must recognise a person it has already counted within a time window you choose. That step is called deduplication.
Beam and thermal sensors cannot do it: they never see a particular person. Wi-Fi tries to do it by device, which randomised addresses now undermine.
Video analytics can do it in two ways. The lighter way follows each person while they stay in view, so someone who steps back and forth is counted once. The stronger way also recognises a person who leaves the view and returns, by comparing how they look, often including the face, for the length of the counting window.
The window is a business decision. A short window counts someone who returns after lunch as a second visit. A full day window counts them once. Choose the window that matches your question and keep it fixed, so trends stay comparable.
Two refinements matter. A minimum time in view stops people glimpsed at the edge of the frame from counting as visitors. And staff will inflate any count unless you plan for them, for example with a counting line they do not cross.
Across several entrances, the dependable figure for a whole building is occupancy, built from entries minus exits at every door. Unique visitor counts are strongest per entrance.
Placement decides accuracy more than any other factor.
Overhead is best for doorways. A camera looking straight down sees each head separately, so people walking close together stay apart and nobody hides behind anyone else. It also sees fewer faces.
Steep angles work for wider spaces. A steeply angled camera can count a broad entrance or concourse. The shallower the angle, the more people overlap, so keep it as steep as the mounting allows.
Draw the line where people commit. Put the counting line where people walk through, not where they stop to wait or chat.
Cover the full width. A wide entrance may need two cameras or a wider lens. Anyone who walks around the end of a line is never counted.
Mind the light. Sun through glass doors, deep shadows and reflections on polished floors cause most camera counting errors. Check the view at the brightest and darkest times of day.
Count every entrance. Occupancy is only as good as the least watched door. Include staff entrances and car park lifts.
People counter accuracy depends on the method, the placement, the lighting and how busy the door is, so a figure measured at someone else's entrance tells you little about yours. The only number worth trusting is the one you measure on your own site.
Ask any vendor who quotes a percentage how it was measured, at which doors and over how many people. Then test it yourself.
1. Pick the hardest door. Choose your busiest entrance at its busiest time.
2. Record the video. Capture at least an hour of footage covering the counting line, including a rush.
3. Hand count it. Count ins and outs from the recording, pausing as needed. Two people counting separately give a reliable reference.
4. Compare like with like. Take the counter's ins and outs for exactly the same period. Check each direction separately, because a miss in one direction can hide a miss in the other when you look only at the net figure.
5. Find the misses. Watch where the two disagree. Groups, trolleys, prams and people lingering on the line explain most gaps, and most are fixed by moving the line or the camera.
6. Repeat at night, in bright sun and on a peak day.
Keep the footage and hand count as your baseline.
People counting can be done without knowing who anyone is. Whether a given system is anonymous depends on what it processes and keeps.
The European Data Protection Board draws the line clearly in its guidelines on video devices. It describes simple counting algorithms as less intrusive than biometric technologies. It also says that a system using physical characteristics to detect a person coming back into view is a biometric identification method, which carries the strictest rules under GDPR. In practice, counting crossings is light touch, while recognising a returning visitor by their face is not.
In the UK, the Information Commissioner's Office guidance on video surveillance expects a data protection impact assessment before processing begins, and states that this is a legal requirement in most video surveillance cases. In India, the Digital Personal Data Protection Act, 2023 governs personal data in digital form, which includes video of people who can be identified.
Good practice looks the same everywhere. Use the lightest method that answers the question: occupancy and passes rarely need anyone to be recognised. Store totals, not images or identifiers, and hold any matching data only for the counting window. Never extract age or gender you do not need. And count on your own servers, so footage of your visitors never goes to an outside provider.
This is general information, not legal advice. Take advice on the rules that apply to your sites.
Retail. Conversion rate, transactions divided by unique visitors, separates a quiet day from a selling problem. Counts also set staffing by real traffic.
Offices and campuses. Occupancy by floor shows which spaces are used and when, shaping desk policy, cleaning and energy use.
Transport hubs. Counts at gates and platforms show peaks as they build, so staff can open another gate before a platform overfills.
Venues, malls and places of worship. A live occupancy figure and an alert when an area passes its limit turn a safe capacity into something staff can act on.
Queues. Counting the people in a queue zone gives queue length and wait time, and an alert on a long line opens another counter before customers walk out.
When not to bother. For one rough weekly figure at a small door, a beam counter may be enough. And if no one will act on the numbers, precise counting changes nothing.
Visylix is an automated AI video surveillance platform built in India. It counts people on the IP cameras you already own, inside the platform that runs your security, so the camera that watches an entrance also counts it. People counting draws on several of its 22 AI analytics, all built in house.
Directional line counting. Draw virtual lines for people or vehicles and get in, out and net counts per line, with hourly breakdowns on a dashboard. Counts hold steady when someone lingers on the line, each crossing counts once, and lines can run on a schedule.
Live occupancy across every door. Several cameras and entrances feed one shared occupancy count for a building or floor, entries minus exits. It resets daily at a local time you choose, never drops below zero, keeps yesterday's totals, and an operator can correct it by hand.
Capacity alerts. Set a maximum occupancy for any camera or zone, and how long it must be exceeded, and Visylix raises one alert per breach rather than a flood.
Unique visitors, counted once. Unique person counting counts each person once per visit on each camera, only after a short minimum time in view, and someone who steps away and returns within the window you set is still counted once. Matching data is held in memory only for the counting window and then discarded. Age and gender are never extracted. Where you prefer not to process faces at all, a tracking only mode counts without them.
Crowds and queues. Crowd counting counts every person it detects, zone by zone, even small, distant people in large crowds. Unless you set your own thresholds, crowd levels from low to critical calibrate themselves to each camera's normal; density per square metre is available on a calibrated camera, and a crowd forming alert fires when headcount rises faster than usual. Queue analytics report live queue length, average and maximum wait, how fast the line is clearing, and a predicted wait for someone joining now, with alerts on length or wait. Movement heat maps show where people go.
Counts that act. With Sudarshan, the Visylix automation layer, a line crossing in a chosen direction, a crowd threshold, a crowd forming or a queue threshold can notify the right team, raise an alarm incident in the operator console, send signed webhooks with automatic retries, publish over MQTT, forward to your SIEM (security event monitoring) or PSIM, and start recording. Every count is also available through the REST API, ready to join with your sales or building systems.
Your data stays with you. Visylix ships as a Docker image that runs on your own hardware: on premise, at the edge, hybrid or fully air gapped. Video never leaves the building.
Plans. Starter and Pro include unlimited cameras on the VMS licence, from ₹4,999 or $49 per month. AI analytics, including people counting, are a separate subscription billed per camera per month, so you switch counting on only at the entrances and zones you choose. Talk to us for AI pricing.
Proven on your doors. We run Visylix on your own entrances during a proof of concept, and you check it against your own hand count.
People counting earns its place at store and mall entrances where conversion matters; in offices planning space on real occupancy; at transport hubs, venues and places of worship that must stay within safe capacity; at checkouts where queues cost sales; and anywhere existing cameras can count without new hardware.
It disappoints when the camera or sensor is badly placed; when one entrance is left uncounted; when passes are reported as visitors; when nobody tests the figures against a hand count; when face based matching is switched on where a lighter method would answer the question; and when the numbers never lead to an action.
A people counter detects each person at a fixed point, usually a doorway, and records the crossing with a direction and a time. Infrared beam counters count breaks in a beam. Thermal and 3D counters look down from the ceiling. Wi-Fi counters listen for phones. AI video analytics detect people in camera footage and count them as they cross a virtual line or enter a zone.
Accuracy depends on the method, the mounting position, the lighting and how busy the entrance is, so a figure measured at someone else's door tells you little about yours. Test any counter yourself: record an hour of footage at your busiest door, hand count the ins and outs, and compare each direction separately with what the counter reported for the same period.
It depends on the job. A dedicated 3D overhead sensor is excellent at one well defined doorway but does nothing else. A people counting camera with AI video analytics can count several entrances, open spaces and queues, raise crowd alerts and keep recording for security, often on cameras you already own. Many organisations choose cameras for breadth and sensors for a single critical door.
It can be fully anonymous: counting crossings or occupancy needs no one to be recognised. Counting unique visitors is different, because recognising a returning person may compare how they look, sometimes including the face. Ask any vendor what is processed, where, and for how long. Good systems keep only counts, hold any matching data briefly, and offer a mode that skips faces entirely.
The system must recognise that a person crossing again has already been counted within a time window you choose, which is called deduplication. Video analytics follow each person while in view and can also recognise someone who leaves and returns within the window. Beam and thermal counters cannot do this, and Wi-Fi counting is weakened because phones now use randomised addresses.
Footfall is how many people came through over a period, measured either as passes or as unique visitors. Occupancy is how many people are inside right now, calculated as entries minus exits across every door. Footfall suits sales and marketing analysis. Occupancy suits safety limits, cleaning, energy use and staffing decisions made in the moment.
Usually, yes. AI video analytics can count people on standard IP cameras if the camera has a clear view of the entrance, a reasonably steep angle, enough light and covers the full width of the door. Some existing cameras may need repositioning or a second camera on a wide entrance. Test the view against a hand count before relying on the figures.
The best people counting software is the one that produces the number you need, at your own doors, with figures you have checked against a hand count. Look for directional in and out counts, live occupancy across several entrances, unique visitor counting with clear privacy controls, capacity alerts, an open API, and the option to run everything on your own servers.
Yes. Visylix counts people on your existing IP cameras with directional in, out and net counts, live occupancy shared across several doors, maximum occupancy alerts, unique visitor counting per camera, crowd levels and queue wait times. Counts can trigger alerts, webhooks, MQTT and SIEM events, and everything runs on your own servers. AI analytics are billed per camera per month on the cameras you choose.