Glossary

Biometric categorisation — inferring characteristics from biometric data, as opposed to identifying someone. Inferring sensitive attributes this way is prohibited in the EU.

Biometric template — a mathematical representation of a face, gait or other physical characteristic, computed for comparison. It is biometric data from the moment it exists, whether or not an image is retained.

Detection — determining that an object of a given class is present. No identity involved, and the tier most operational questions actually need.

Drift — degradation of performance over time from changes to the camera, the scene, the lighting or the population, with no alert.

Ground truth — an independent record of what actually happened, collected without seeing the system's output. The only basis for measuring accuracy.

Identification — determining which specific known person this is, by searching a gallery. One-to-many, uncooperative, and the tier with the heaviest obligations.

Operating point — the threshold at which a score becomes a decision. A business decision, usually left at a default nobody chose.

Pixels on target — how much sensor resolution falls on the thing being detected. The number that predicts whether detection is possible at all.

Precision — of the alerts raised, the proportion that were real. The alert-fatigue number.

Recall — of the events that occurred, the proportion detected. The safety-relevant number.

Verification — one-to-one confirmation that a person is who they claim, with their participation. A materially different proposition from identification, and routinely conflated with it.

Terms used loosely elsewhere

"Accuracy" as a single percentage conceals the threshold, the base rate, the conditions and the distribution across people. These notes ask for four numbers instead.

"Anonymised" rarely is. A movement record without a name is usually still personal data.

"AI-powered" says a learned model is involved and nothing about capability. For measurable properties a rule is frequently better.

"Behavioural analytics" covers dwell time and intent inference alike. These notes distinguish them, because one is a measurement and the other is a claim about someone's mind.

"Privacy-preserving" sometimes means frames are discarded at the edge, which is genuine, and sometimes means faces are blurred in the interface while templates are stored, which is not.

"Compliant" describes a deployment, never a product. A vendor cannot make your deployment lawful.

On the absence of numbers

A note on why this collection quotes so few figures.

No accuracy benchmarks are given, deliberately. Performance depends on cameras, lighting, scene, population and threshold, and a number from another deployment predicts nothing about yours.

No industry error rates are quoted for the same reason, and because published figures are self-selected.

Where a relationship is stated — that error rates fall unevenly across groups, that precision collapses for rare events, that alert volume determines whether anyone reads alerts — that is the part worth relying on.

Every number worth having is your own, measured on your site, twice.

How to read these notes

A short orientation for anyone arriving at the collection rather than at one page.

Start with the foundations if the subject is new, particularly the note separating detection from identification.

Start with the boundary section if something in that tier has been proposed, before anything technical.

Start with the limits section if a demonstration was impressive.

Start with getting started if you have cameras and a question.

Start with common failures if a deployment is disappointing.

The reference section states what the system cannot do, which requests to refuse, and what a working arrangement looks like.