To exclude staff from visitor counts, a video analytics solution must treat employees as a separate category and then exclude them from foot traffic metrics. This is done without facial recognition: the algorithm does not seek to identify a person; rather, it classifies a silhouette based on stable visual cues, such as work attire, and on contextual rules, such as entering an area reserved for staff.
Several French video analytics providers offer this feature, including XXII with its CORE platform. The key consideration is not the existence of the feature itself, but its reliability under real-world conditions, how to test it, and the legal framework that must be followed whenever employees are being filmed. The following sections detail the available methods, their limitations, and the checks that must be performed.
A foot-traffic counter does not automatically distinguish between a customer and an employee. Entering a store, passing in front of an aisle camera, or walking through a security gate are all counted the same way. In a retail location where staff move around frequently—entering and exiting for the backroom, deliveries, or breaks—the proportion of non-purchasing visits becomes significant.
This effect extends to all metrics calculated based on foot traffic counts. The conversion rate—which compares the number of sales receipts to the number of visitors—is underestimated. The average dwell time is skewed upward or downward depending on staff habits. Comparisons between stores become unreliable, since two stores with identical actual foot traffic may show discrepancies that merely reflect differences in staffing levels or organizational structures.
The problem, therefore, has less to do with absolute accuracy than with comparability. A constant, known error can be corrected. An error that varies by location, time, and day makes operational decisions unreliable.
Facial recognition involves comparing a biometric template to a reference database to identify a person. This processing falls under a specific category of data and requires a strict legal framework. Staff exclusion does not require this, as it never requires knowing who the person is, only whether they belong to the group of employees.
The approaches used rely on three types of signals.
The model learns to recognize a visual attribute associated with work attire: a vest, an apron, a polo shirt of a specific color, or a badge worn prominently. This approach works well when uniforms are consistent and high-contrast. Its effectiveness decreases when staff wear personal clothing, when a customer wears a similar color, or when lighting affects color perception.
A person who exits the stockroom, walks through a service door, or stands for an extended period behind a register is highly likely to be a staff member. These zone-based rules are robust and inexpensive to configure. However, they require that the camera actually cover the relevant access points.
Movement patterns differ: an employee makes repeated trips back and forth along the same path, stops at fixed locations, and crosses areas that a customer does not. When analyzed over time, this movement signature usefully complements the first two signals.
In practice, the most stable solutions combine all three categories rather than relying on just one.
No provider can guarantee perfect exclusion. The right question, therefore, is not about reliability in absolute terms, but about the level of error that remains in your store, given your security protocols and cameras.
A useful verification process involves a few steps:
Requesting an overall error rate without this level of detail can lead to an unpleasant surprise: a solution may show an excellent average while incorrectly misclassifying some customers during peak hours—precisely when the measurement matters most.
As soon as a video system processes images of employees, the GDPR applies, and several obligations are added to those under labor law. The purpose must be defined and limited: measuring customer traffic does not authorize the monitoring of an employee’s individual activities. The principle of data minimization requires retaining only necessary data, and counting generally produces aggregated data rather than stored images.
It is mandatory to inform individuals—both employees and customers. Furthermore, the implementation of a system capable of monitoring employee activity requires prior consultation with employee representatives. The CNIL publishes recommendations on video surveillance in the workplace and on systems for measuring foot traffic in public spaces: consulting these guidelines before deployment prevents having to revise a project after the fact.
One point deserves clarification in the project documentation: the exclusion of staff is a statistical filtering operation, not individual tracking. Framed this way, it is easier to justify to employee representative bodies.
Once staff are excluded, three metrics become actionable. The conversion rate finally reflects the ratio of visitors to buyers and can serve as a basis for comparisons between stores or over different time periods. Staffing levels are determined based on actual customer foot traffic trends rather than a mix of the two populations. Finally, the evaluation of a sales promotion or a layout change is based on a less distorted measurement.
The magnitude of the gain depends entirely on the context: floor space, staff size, and the organization of stock-filling flows. It is measured on a site-by-site basis, which is why the manual control count described above remains essential before rolling out the system on a broader scale.
XXII is a French computer vision software company. Its CORE platform analyzes foot traffic data from existing cameras and generates operational metrics without using facial recognition. For a retail chain, the main benefit lies in the consistency of the measurement: the same definition of a visitor, applied uniformly across the entire network, makes comparisons between locations reliable.
That said, decisions are based on evidence. Before proceeding with a deployment, request a cross-validation analysis using your own cameras, and focus the discussion on cases where the system fails to distinguish correctly rather than on cases where it succeeds.