People counting is the analytic that tallies how many people enter, leave, or occupy a space — the foundation of retail and facility analytics. A camera over an entrance counts footfall; cameras over a zone estimate occupancy. The output feeds conversion metrics (visitors versus sales), staffing decisions, queue management, and occupancy-limit compliance, turning a camera into a sensor that measures demand rather than just records it.
Its accuracy should be stated as a range, not a single figure: well-placed people-counting commonly runs in the 90–98% range, and reputable vendors decline to promise a single number because it depends on camera angle, height, lighting, and crowd density. An overhead view counts far better than a shallow one, and the IEC 62676-4 DORI framework is a reminder that the counting camera (which needs only to detect a body, roughly 25 px/m) is not the identifying camera (roughly 250 px/m) — different jobs, different placement. Counting can run at the edge, sending only the tallies as metadata over ONVIF Profile M and cutting bandwidth dramatically.
The privacy point is a genuine advantage here and a pitfall if crossed. Plain people counting is anonymous — it counts bodies, not identities — so it generally sits outside GDPR Article 9, and the EDPB Guidelines 3/2019 treat simple counting as less intrusive. But bolt on face matching or a watch-list and it becomes biometric processing under Article 9 and Illinois BIPA, a completely different legal tier (the ICO actions against Serco and around Facewatch illustrate the line). Keep counting anonymous unless there is a justified, lawful basis to identify, and never quote the count as 100% accurate. This is engineering guidance, not legal advice.
