Occupancy and Access Logging From Cameras You Already Own

Video analytics is usually sold as a list of features, which hides the fact that they differ enormously in how dependable they are. Below is the same list, graded — because knowing which capability will disappoint you is more useful than knowing what is on the datasheet.

Graded by how much you can lean on it

Why occupancy drifts and counting does not

A count is an event: someone crossed a line, the number went up by one. An error affects that one event. Occupancy is a running total of entries minus exits, which means every error is permanent and they accumulate in one direction all day. Two missed exits an hour is an invisible error on a count and a badly wrong occupancy figure by mid-afternoon.

This is why any occupancy system worth using resets — overnight, or on a known-empty condition — and why a vendor who cannot tell you the reset behaviour is describing a number that will be wrong by Thursday. Ask what happens when two people walk through shoulder to shoulder, and what happens when someone stands in the doorway holding it open. Those are the two events that generate most of the drift.

Access logging without identification

The most useful and least oversold capability in the list is the plain one: a timestamped record that someone entered a restricted area, with the clip attached. It does not say who. For a great many purposes that is sufficient — a plant room that should have had nobody in it at 02:00 is an actionable finding regardless of identity, and the clip lets a human answer the identity question in seconds.

It is worth resisting the upgrade to named identification unless you genuinely need it, because that changes the legal category rather than just the feature set. Detecting a person is not biometric processing; recognising a specific individual is, and under Article 5 of the PDPL that requires a Ministry permit before you begin — the same constraint attendance systems face. Anonymous zone logging keeps you out of that entirely while answering most operational questions.

Camera position decides more than model quality

Almost every disappointing deployment traces back to geometry rather than software. Counting wants a camera looking down at a doorway, close to overhead, where two people side by side are two distinct shapes. Most existing CCTV is mounted high on a wall looking across a room, because it was installed so a human could recognise faces — which is the opposite optimisation. At that angle people occlude each other and a count degrades in a way no model fixes.

So the honest version of "use the cameras you already have" is: some of them, for some tasks. Zone presence tolerates almost any viewpoint and works on the existing estate. Counting and occupancy often need one camera repositioned or added at the door that matters. Budgeting for a small number of repositions is more realistic than expecting an estate installed for human viewing to serve analytics unchanged.