Camera-based people counting in retail: methods, accuracy and limits

Why count visitors?

Revenue and transaction counts tell you what was sold, not how many people came in. Without footfall data, it is hard to tell whether a drop in sales is down to less traffic or to difficulty turning visitors into buyers.

Camera-based counting lets you measure entries hour by hour, compare days and stores, match staff rotas to actual peaks and assess the impact of a window display or a promotion. That only works if the figures are reliable and correctly interpreted.

For a network of several stores, there is an additional challenge: comparability. Two stores can only be compared if their entries are measured with the same method and with accuracy checked at each one. An isolated figure, however precise, is worth less than a consistent series over time and across sites.

The main counting methods

Two measurement approaches are used, often together.

Line-crossing counting draws a virtual line, usually at the entrance. Each person who is detected and tracked crossing that line is counted as entering or leaving depending on their direction. This is the standard method for measuring store footfall.

Zone counting measures how many people are in a defined area at a given moment: an aisle, a department, the checkout area. It is used to estimate occupancy, queue build-up or how much an area attracts shoppers, rather than total visitor numbers.

Camera position matters just as much:

• Overhead (top-down) camera: mounted on the ceiling above the entrance, it greatly reduces occlusion between people, because each person is seen separately. It is generally the most reliable set-up for counting entries, but it needs enough ceiling height and a dedicated installation.
• Angled camera: this is how most existing security cameras are mounted. It covers a wider area, but people can hide one another, especially in groups or at peak times.

What affects accuracy

Be wary of accuracy rates quoted without context: how reliable a count is depends above all on the conditions on site. The main factors are:

• camera angle and height, and therefore the degree of occlusion;
• crowd density: tight groups, pushchairs and trolleys make it harder to separate individuals;
• lighting: backlighting from a glass door, differences between daytime and evening, reflections on the floor;
• the width of the entrance and whether several entrances need covering;
• stream quality and frame rate;
• unwanted crossings: staff, delivery drivers, children, people who hesitate at the door and walk straight back out.

Validating the figures during a pilot

The only serious way to know how accurate counting is on your site is to measure it. Before you start, set an acceptance threshold that suits the intended use: adjusting staff rotas tolerates more imprecision than an indicator used to assess a store’s performance. Then express the difference as a percentage of the reference count, slot by slot and cumulatively, without relying on the total alone: missed entries and double counts can cancel each other out and hide a problem. During a pilot, we use manual sample counts:

• choose several representative time slots: a quiet hour, a peak hour, the end of the day, a weekend day;
• count entries manually during those slots, live or from recordings, ideally with two people to reduce human error;
• compare these reference counts with the system’s figures, slot by slot;
• analyse the differences: are they consistent, or concentrated in certain situations (groups, backlighting)?
• adjust the framing, counting line or settings, then run another round of checks.

Video-measured visits and sales are not the same thing

A video indicator measures physical movements. The conversion rate compares that visitor number with the number of till receipts over the same period. Calculating it requires integration with the point-of-sale (POS) system, which depends on its export or API capabilities.

A few interpretation caveats apply: a group or family generates several entries for a single receipt, staff should be excluded as far as possible, and time windows must be aligned. Conversion rate is therefore most useful as a trend, for comparing periods or stores measured in the same way, rather than as an absolute value.

In a well-designed dashboard, every indicator clearly shows its source: video for entries, the till for sales, and a combination of both for conversion rate.

Counting and privacy

Counting people does not require identifying them. A privacy-friendly solution simply detects human shapes, produces aggregated counts and keeps only those results, with no facial recognition and no individual profiles.

On-site processing (Edge AI) avoids moving footage around, and masking zones exclude irrelevant parts of the image. The retailer, as data controller, must also inform customers and staff with appropriate signage and complete the applicable formalities with the CNDP.

It is worth documenting these choices from the pilot stage onwards: which cameras are used, which areas are masked, what data is kept and for how long, and who can access the dashboards. This documentation supports your regulatory formalities and helps you answer questions from customers or staff clearly.

In summary

Camera-based people counting can become a genuine management tool, provided you choose the right method, position the camera correctly, measure accuracy on site rather than assume it, and keep video data clearly separate from sales data.

Validation is not a one-off exercise. A camera that has been knocked slightly, a new display stand near the entrance or a change in lighting can degrade a count that used to be reliable. Plan periodic sample checks and an alert when a camera stops sending data or reports abnormal values. Thinking about measuring footfall in your stores? Start with an audit of your entrances and a well-scoped pilot.