Alternatives to facial recognition on CCTV

Facial recognition promises identification and delivers three problems instead. It's the highest-risk category of processing under UK GDPR, the public visibly doesn't trust it, and on real CCTV the face is often at the wrong angle, too far away, or under a hood anyway.

Why operators are stepping back from faces

The regulatory reason is straightforward: biometric identification is special category data under UK GDPR, carries heavy obligations, and sits under active attention from the ICO. The trust reason is just as real, customers and staff react badly to being face-scanned, even when the intent is reasonable. And the practical reason is the one most operators discover for themselves: cameras are mounted high and wide for coverage, not for faces, so a usable face capture is the exception on real footage, not the rule.

What works instead: describe, don't identify

A vision model describes every kept frame in plain English: the people in it, what they're wearing, the objects around them, what's happening. Those descriptions are what gets searched. "Grey coat, black cap, heavy build, near the loading door" finds the moments that match. Nothing biometric gets created in the process, and nobody is identified or enrolled. This stops deliberately short of re-identification, the industry term for matching a person across cameras by something like their outfit or gait, which Svid doesn't implement.

What you give up, honestly

Description search doesn't tell you who someone is, and it doesn't automatically connect their appearances across cameras. Svid deliberately does no re-identification, so if a person moves from the shop floor to the car park, you run the same search on the next camera's footage and confirm the match yourself from the clips. A full change of clothes changes the description, too. For most operational jobs, find the moment, review the incident, hand the clip on, that trade is fine, because knowing a name was never the operational need.

1. 01

   Describe the person or the moment in plain English

   Clothing, build, what they were doing, roughly where and when. That's enough to search on.

2. 02

   Jump straight to the matching moments

   Each result links into the footage at that point. No scrubbing required.

3. 03

   Run the same description over other cameras, if the story spans more than one

   The full workflow for finding someone in footage covers searching across a folder. A human confirms any match between cameras; Svid doesn't connect them automatically.

4. 04

   Export or share the moments you need

   At no point was anyone biometrically identified or tracked.

Can you find someone on CCTV without facial recognition?

Yes, by searching a plain-English description of what's visible: clothing, build, what they were doing. That's the whole approach in a sentence.

Is description search legal under UK GDPR?

It avoids the biometric high-risk category and creates no identity data. Usual operator duties still apply (signage, lawful basis, retention policy); this isn't a compliance guarantee or legal advice.

Can it follow a person from camera to camera?

No, deliberately. There's no re-identification or people tracking in Svid. You search each camera for the same description and confirm the match yourself.

Why not just use facial recognition anyway?

It's the highest-risk processing category under UK GDPR, the public reacts badly to being face-scanned, and on real CCTV a usable face capture is often the exception rather than the rule.