Watchlist matching
Watchlist matching compares every face a camera sees against a defined list of enrolled people and raises an alert only on a candidate match. It is a 1:N search where the gallery is small and specific, and where the operational question is what happens on an alert, not only how the match was scored.
How does a watchlist system work?
- Enrolment: a small gallery is built from images of the people the list covers, each quality-checked before it is accepted.
- Detection: faces are detected in the live stream and filtered for quality before any comparison happens.
- Search: each usable face is compared against every enrolled template, producing scores.
- Alert: a score above the threshold raises a candidate alert with the probe image, the enrolled image and the score.
- Adjudication: an operator confirms or dismisses. Nothing acts on the match automatically.
The quality filter before the search matters as much as the matcher. A stream at thirty frames per second produces many low-quality views of the same person, and comparing all of them multiplies both the compute cost and the number of chances to raise a false alert on one passer-by.
Why does gallery size govern the false alert rate?
Every enrolled template is an independent opportunity for an unrelated person to score above the threshold. Doubling the list roughly doubles the false alerts at a fixed threshold. This is why operational watchlists are kept deliberately short and reviewed for removals, and why a list that grows without pruning degrades on its own.
- Keep the list to people there is a current, documented reason to alert on.
- Set a review date at enrolment, so entries expire rather than accumulate.
- Use the best available enrolment image; a poor enrolment costs accuracy on every search that follows.
- Track the confirmed-to-dismissed alert ratio, which is the metric that tells an operator whether the threshold is right.
What does an alert workflow need?
An alert is a request for a human decision, not a conclusion. A workflow that treats it as a conclusion is where watchlist deployments cause harm: the technical error rate stops being the risk, and the process built on top of it becomes the risk instead.
Both images should be shown side by side with the score, the operator's decision should be recorded with the alert, and dismissed alerts should be retained for audit and threshold tuning rather than deleted.
Frequently asked questions
- What is watchlist matching?
- Watchlist matching compares every face a camera sees against a defined list of enrolled people and raises an alert on a candidate match. It is a 1:N search with a deliberately small gallery, and the alert is a request for human confirmation rather than an automatic decision.
- How large should a watchlist be?
- As small as the purpose allows. Every enrolled entry is another chance for an unrelated person to score above the threshold, so false alerts scale with the list. Operational practice is to enrol only people there is a current documented reason to alert on, and to set an expiry at enrolment.
- Is watchlist matching verification or identification?
- Identification. The system searches a gallery to determine whether a passing face belongs to anyone enrolled, without the subject making any identity claim. That places it in the more tightly regulated category in most jurisdictions, including under the EU AI Act.
- Should a watchlist alert trigger an automatic action?
- No. A candidate match is evidence for a human decision, and systems that act on the match directly convert a scoring error into a real-world consequence. Standard practice is to display the probe and enrolled images with the score and to record the operator's decision alongside the alert.
- What image quality does enrolment need?
- The best available, because enrolment quality affects every subsequent search against that entry. A frontal, well-lit image with enough pixels across the face is worth more than any threshold adjustment, and a poor enrolment cannot be compensated for later in the pipeline.
- Can watchlist matching run on existing CCTV?
- Often, subject to the pixels across a face at the point where people pass, the mounting angle and the lighting. Cameras positioned for scene coverage frequently do not meet the resolution a search needs, which is why a site survey precedes procurement rather than following it.