People Counting Camera: Accuracy, Cost and Privacy

A people counting camera answers a question no incident report ever captures: how many people actually used this space, and when. It is the least glamorous job in video analytics – no alarms fire, nothing gets intercepted – and yet it is often the one that pays for the system, because staffing, layout and capacity decisions all get better the moment you can see demand instead of guessing it.

What follows covers where the counting is genuinely useful, how it differs from identity-based analytics, where accuracy breaks down, and how the privacy question is answered when you deliberately stop short of identifying anyone.

What People Counting Is Actually For

The value shows up in five recurring shapes:

• Conversion and dwell – compare the number who entered with the number who transacted, then work out which parts of the space hold attention.
• Staffing against real demand – tills, hosts and security posts matched to footfall by hour rather than to a fixed rota. The mirror image of this problem is covered in under-staffing detection.
• Capacity and safety – knowing a hall, platform or staircase is approaching its occupancy limit before it becomes a crowd problem, which is the territory of overcapacity detection.
• Space utilisation – which entrances, corridors and zones are used, and which were designed on an assumption that never held.
• Evidence for planning – a year of hourly counts is the argument that wins a refit budget, and it is data you cannot reconstruct after the fact.

Counting Is Not Recognising

This distinction matters more than any accuracy figure. Counting requires the system to notice that a person exists, track them as an object through the frame, and increment a number. It does not require knowing who they are. Face recognition does the opposite – it is built on identity, and it carries a different set of legal and ethical obligations.

That is why counting is usually the easier conversation with a privacy officer. There is no gallery of faces, no match event, no biometric template to store. Where a site does need identity, that is a separate project with its own justification – see face recognition and verification for what that involves.

If the underlying idea of turning video into numbers is new to you, what AI video analytics does is the right starting point.

How the Counting Works

Under the hood it is three steps repeated per frame. The model detects people and gives each one a temporary tracking identity. As a tracked person crosses a line drawn in software – a doorway, a lane, a zone boundary – the tally increments or decrements depending on direction. Because identity persists across frames, the same person is not re-counted on their next step, which is exactly where naive motion-based counting fails.

Two configuration choices do most of the work. A crossing line placed in a doorway gives you entries and exits. A zone gives you a live count of who is currently inside, which is what feeds occupancy rules. Most sites end up using both: lines at the entrances, zones over the areas where a number matters. Where a count needs to reflect how long people linger rather than how many pass through, the dwell-time rules in person dwell detection cover that case.

Where Accuracy Breaks Down

Honest expectations beat marketing numbers here. Five conditions cause most of the error, and all five are visible on a site survey:

• Density. Once people are shoulder to shoulder, bodies merge into one blob and any algorithm undercounts. A counting system is reliable in flowing traffic, and progressively less so as a space saturates.
• Occlusion. Columns, shelving and signage hide people from the line of sight. A camera that sees the floor but not the gaps between aisles is counting a fraction of the room.
• Line placement. A crossing line where people pause – a till, a lift lobby, a stair landing – produces double counts as they step back and forth. Put lines where movement is continuous.
• Lighting. A dark entrance or a night-time car park degrades any video model. That is a lighting and camera problem, not an algorithm problem.
• Reflections and screens. Glass facades and mirrors can create duplicate-looking figures.

None of this means counting is unreliable – it means the accuracy of a counting system is decided at the site survey, not at the model. Budget a commissioning period too: review the first week against a manual spot count, then tune. Sites that skip that step tend to blame the algorithm for what is a camera-angle problem.

Privacy: Counting Without Identifying

Because no identity is required, a counting deployment can be designed to stay clear of personal data altogether. Three design choices carry most of the weight, and they are the ones worth putting in writing before go-live:

Process on site, not in a cloud. When analytics run on an appliance in the building, footage is not uploaded to a third party for analysis. The footage that never leaves the premises is footage that never enters someone else’s data-processing agreement – the trade-offs are set out in on-premise versus cloud analytics.

Keep the output statistical. What the business needs is a number, a trend and a timestamp – not a roster. Configuring the system to emit counts and events rather than searchable face galleries is what keeps the deployment on the right side of the line.

Agree retention deliberately. Counts can be kept for years; raw video should not be. A written retention window, and a named list of who can review footage, are the two questions any auditor will ask. For the privacy framework these answers map onto, see the NIST Privacy Framework.

Running It on the Cameras You Already Have

Counting does not need new cameras – it needs a reasonable view. An edge AI box ingests the streams your existing units already produce, runs the models on-device, and holds the counters as configuration rather than hardware. Systems scale from 2 to 128 channels with 1 to 256 TOPS matched to load, carrying a library of 198+ pre-built algorithms in software. Practically, that means the entrance camera doing door duty can also be the one that counts, and adding a queue rule later is a settings change rather than a purchase order.

Retail sites usually start here because the numbers connect directly to money – see AI for retail and retail loss prevention for how counting sits alongside the security detections on the same cameras.

What It Costs

Adding analytics to existing cameras starts at USD 399 for Standard 2, with Standard 4 at USD 999 and Standard 6 at USD 1,599; larger channel counts are quoted per site. Each Starter Kit ships with 10 metres of cabling as standard across the whole kit, with additional cable at USD 25 per 10 metres. The full breakdown is on our pricing page.

For a configuration-led project, the first working demo on your own footage typically lands in about seven days. Bespoke algorithm development is different: until we have seen your video – the angles, the lighting, the specific thing you need counted – nobody can honestly commit to a date, and we do not.

This is one of the topics we cover in more detail in smart city video analytics.

For a site-by-site breakdown of running AI on a live build, see our guide to Construction Site Security Cameras with On-Site AI.

Frequently Asked Questions

Does people counting use face recognition?

No, and it should not need to. Counting tracks people as anonymous objects and reports totals – it never establishes who anyone is. Keeping those two capabilities in separate projects is both a privacy decision and a simpler compliance story.

How accurate is it in a busy space?

Accuracy is governed by density, camera angle and lighting rather than by the model alone. Flowing traffic through a doorway is the easiest case and the most reliable. Once a space becomes shoulder to shoulder, all video-based counting undercounts, and a better camera position helps more than any parameter change.

Can I count without storing any video at all?

Yes. Some sites run analytics purely for their numbers and keep no archive. Others keep footage for a defined retention window and store only counts beyond it. Which you choose is a policy decision, and it is worth writing down before the system goes live rather than after.

Will it work with the cameras we installed years ago?

Usually, provided the camera delivers a standards-based stream and has a view of the area you care about. Age matters far less than stream compatibility and angle. During evaluation each camera is checked against the rules you want to run, and unsuitable positions are flagged before you commit.

Next Step

Send us a floor plan or a photo of the entrances and the areas you want counted, and we will tell you which of your existing cameras can carry the rule and what the configuration would look like. For context on occupancy and life-safety thresholds, see the National Fire Protection Association. Start with a Starter Kit to validate counting on your busiest entrance, or tell us your scenario for a configuration quoted per site.