How to find a person in CCTV footage

You usually know two things: roughly when, and roughly what they looked like. An incident on Tuesday afternoon, a contractor in hi-vis who was near the loading dock, a customer in a red jacket who left something behind. The traditional method is scrubbing at eight times speed and hoping your attention holds, and on a multi-camera site the maths never works. Eight cameras over one afternoon is a full working day of watching.

Describe them, don't scrub for them

When footage comes into Svid, a vision model reads it frame by frame and writes down, in plain English, who's visible and what they're doing. Those descriptions are what you search. Finding a person stops being a viewing job and becomes a sentence: type “man in a grey coat near the spirits” and the matching moments come back in seconds, each stamped with a time and linked straight to the point in the footage. The search bar covers whatever folder of the Library you're standing in, subfolders included, so one search can cover a single camera or a whole building.

1. 01

   Get the footage in

   Upload the exports, paste a video URL, or connect the camera feed. Live and recorded footage go through the same pipeline, so the pile of clips the recorder spat out is a perfectly good starting point.

2. 02

   Search from the right folder

   The Library is a folder tree, and search covers the folder you're in, subfolders included. Searching from the site's folder covers every camera in it; searching from one camera's folder narrows to that view.

3. 03

   Type the description

   Outfit, colours, anything distinctive, plus where they were if you know it. You're writing the sentence you'd use to describe them to a colleague, not filling in a form.

4. 04

   Narrow if you need to

   If the first pass comes back wide, add the place or the time. “Red jacket” across a week is a crowd; “red jacket near the tills on Tuesday afternoon” is usually one person.

5. 05

   Jump to each match and confirm

   Every result links to its moment in the footage, so confirming it's them takes a glance rather than another viewing session.

What makes a good description

Specifics beat suspicion. “Red jacket, black cap, carrying a holdall” finds people; “suspicious man” finds nothing, because the search matches what's visible in the frame, not a mood. Outfit colours and distinctive items are the strongest signals, and a place helps too (“near the spirits”, “by the fire exit”), because the descriptions record where things happened in the scene. If the clothing is common, lean on what they were doing instead: carrying a ladder, or standing at the counter far longer than anyone stands at a counter.

What Svid deliberately doesn't do

Svid finds moments that match a description. It never builds a person. There's no following someone from camera to camera, no identity profile that accumulates sightings, no facial recognition, and no fingerprinting anyone by outfit, gait or height. That's a design decision rather than a missing feature: systems that re-identify people sit in the highest-risk category of processing under UK GDPR guidance, and the ICO expects a great deal of an operator before that kind of surveillance is lawful. A business looking for one person on one afternoon has no appetite for any of that. Describing footage and searching the descriptions answers the question without putting a tracking system on your compliance register.

If you need the route, not just the moment

Sometimes the question isn't “where were they at 4pm” but “how did they move through the site”. You answer that with a sequence of searches: run the same description over each camera's footage, note the times, and the order falls out on its own. If you need it in writing, run the Incident Reconstruction report check over the relevant videos; report checks read a video's whole timeline and write a document, so you get the sequence as an ordered account rather than a set of open tabs. It's assembled from what's visible in each frame, not from tracking anyone.

Can I find someone on CCTV without facial recognition?

Yes. You search the plain-English descriptions Svid wrote when it read the footage, so “woman in a green coat with a pushchair” finds the moments she's visible. There are no face templates or biometric IDs anywhere in the process.

What if I only know what they were wearing?

That's the ideal input. Outfit descriptions are exactly what the search matches on, so “blue puffer jacket and white trainers” is a strong query all by itself.

How long does it take to search hours of footage for a person?

Seconds. The search runs over stored descriptions rather than the video itself, so hours of footage answer about as fast as a single clip would.

Can Svid follow a person from one camera to the next?

No, deliberately. Each search finds moments matching your description; nothing links them to an identity. To establish a route, run the search over each camera's footage and put the timestamped moments in order, or have the Incident Reconstruction report write the sequence up for you.

Does it work on old exported footage or only live cameras?

Both. Live and recorded footage go through the same pipeline, so you can upload an export from last month and search it exactly the same way.