Capability
CCTV face recognition — the difference between 'someone is there' and 'who is there'
CCTV face recognition compares a captured face with an authorized enrolled set. A possible match can support review, but it does not establish permission to enter or whether a situation is safe. Someone familiar may still enter a restricted area; someone unmatched may be an enrolled person whose capture was unsuitable. Assess the camera view, matching conditions and alert rules at your site, and confirm the supported workflow in the written scope.
How cctv face recognition works
Enrolment
Use the selected system's supported enrolment procedure for authorized participants. Confirm sample requirements, identity mapping, permission and retention before capture; do not assume existing footage is an approved enrolment source.
Template, not photograph
Recognition systems convert a face into a mathematical vector for matching. Whether enrolment images are kept or discarded, and for how long, is set out in the written scope and your retention policy.
Matching at the edge
Faces the cameras see are compared against the enrolled set. Processing runs on a unit at your site, and how matching is configured for your site is confirmed in the written scope.
Act on the result
Rules can be configured and tested at your site — for example, silence for a known face during expected hours, an alert with a snapshot for an unknown face at a sensitive door, and an alert for a known face somewhere they shouldn't be.
Where it earns its keep
- Assessing factory gate attendance where finger-based capture is difficult, with a supported fallback and review of missed records
- Silent entry for family in a home system, so notifications stay worth reading
- Reviewing unmatched-person events at school gates within an authorized, documented safeguarding workflow
- Restricted-room access logs in hospitals and offices
- Reviewing relevant observations of authorized enrolled participants in retail areas, within the agreed scope
What it won’t do
Every AI vendor lists strengths. These are the limits, so you can plan around them instead of discovering them.
- A turned or covered face, backlighting and occlusion can affect capture or matching. A person-detection event is not guaranteed when identity matching is unavailable; test the actual conditions and define the review or fallback.
- Recognition quality depends on the actual capture view, lighting, movement, enrolment and matching settings. Assess the proposed camera and system together; a mounting height alone does not establish accuracy.
- It identifies enrolled people. It cannot tell you the name of someone who has never been enrolled, and no honest system claims otherwise.
Face recognition — common questions
Can face recognition work on my existing CCTV cameras?
Sometimes. Assess usable face detail, capture angle, lighting, movement, stream compatibility and the selected software together. Neither the megapixel label nor a fixed mounting height establishes suitability. Test representative passages on the proposed camera view and confirm any additional equipment or supported fallback in the scope.
Are photos of my staff stored somewhere?
What is stored, where and for how long is agreed in the written scope and your retention policy before anyone is enrolled. Processing runs on a unit at your premises rather than in a cloud.
How accurate is it?
PGAK has published no accuracy figure for face recognition. Agree an authorized trial that records correct matches, incorrect matches, missed matches and capture failures under representative conditions. Angle, lighting, movement, face coverings, enrolment and configuration can affect results; measure the proposed system rather than assuming one factor determines accuracy. Document human review and the supported fallback for unresolved cases.
Does it work with masks or helmets?
Results depend on the selected system and actual conditions. Test permitted headwear and required PPE as worn; required protection stays in place. If identity matching is unsuitable, agree a supported alternative and human review rather than treating a missed match as an absence.
Sectors that lean on this: Retail shops and showrooms, Jewellery showrooms, Hotels and restaurants, Offices and IT parks, Schools and campuses, Hospitals and clinics, Housing societies and residential — see every industry PGAK serves.
Calculator · C25
Face recognition and face search: estimate it for your site
Use your own measurements. The result is the staff time released, shown separately from any cash, using the formula below. It runs in your browser and nothing is sent anywhere.
Result
Enter your own figures, or load the worked example, to see the result.
How it is calculated
hours = checks × (old − assisted seconds) ÷ 3600 − false-match review hours
Worked example (illustrative inputs)
1,000 checks; 30 old; 10 new; 1 review hour.
staff time released: 4.556 hours a month.
Formula version 1.0.0.
What this does not tell you
- Hours released are capacity, not cash. They become cash only through a mechanism you state — overtime that stops, a post that is not refilled — entered as the realisation fraction.
- The 'after' values are your assumptions or pilot measurements. PGAK publishes no effectiveness figure for this feature.
- Results must not drive automated punishment, pay deductions, access denial or employment decisions. Use requires appropriate notice, a lawful purpose and human review of every exception.
Evidence status: Supplier/integration or marketed capability; PGAK availability and site performance require verification.
Discuss your camera setup
Share your phone number, city and camera count. We will clarify the intended use case and agree assessment scope and timing with you. Do not submit camera credentials or footage.