Can you find a person without knowing who they are?
Find someone by what they looked like, with no face and no plate
An operator has a description and thousands of hours of video from dozens of cameras. Watching them one by one is not an option, and an ordinary search can only say that something moved there. IRIS notes what every object looked like as it went past: what it was, what colour, which way it went, how long it stayed and at what time. Searching stops being watching video and becomes a question with an answer.
At a glance
What it can do
Search by what it looked like
Fifteen criteria in a single query: object type, colour, match mode and tolerance, behaviour, duration, direction of travel, area of the frame and how it fits, dates, time of day, cameras, number of results, confidence and overlap with motion. They all combine at once, against what was already noted down, so nothing has to be watched again.
One sentence is enough
You type what you are after — red cars between two and six — and the filters fill themselves in. It shows what it understood and what it dropped, and anything the sentence never mentioned is left alone.
Point at it, find it again
Click an object in the picture and look for everywhere else it turns up: that camera, all of them, or a chosen few. Whoever searches sets the minimum likeness, and every hit states its own.
The same one, on another camera
The system proposes possible sightings of the same object on other cameras, ranked by confidence level. It does not decide: a person confirms or rejects each one, and that decision is signed.
The whole route
Confirmed sightings gather into a single file: first and last time seen, how many cameras saw it and where it went. It exports to a spreadsheet and attaches to the incident.
Going back to the video
When what you need was never noted down, the recording itself can be swept again. It says up front what that will cost, shows progress as it runs, and can be cancelled without losing what it already checked.
Who searched for what, and when
Every search leaves a record: who ran it, when, over which cameras and with what criteria. And a user only finds objects on cameras they are allowed to see; the rest never show up, and the screen says so.
Face recognition ships switched off
The product includes face recognition. It ships switched off and is not offered in the European Union. Where the law allows it, turning it on is each site’s own decision, and who did it is on record.
And what it threw out
The panel shows every filter it set — class, colour, time span, cameras, behaviour, zone — and also what it threw out and why: a colour outside the catalogue, a camera you are not allowed to see, a zone that has to be drawn by hand. And when it verifies results, it tells “not checked yet” apart from “does not match”.
Colour, measured the same way every time
Under an orange street light, a white van does not come out white. So colour is worked out from the crop of the object itself, not from the whole frame, and it can be searched three ways: dominant colour only, present anywhere on the object, or by similar shade. Eleven catalogue colours, plus any free colour picked on screen.
What is kept, and for how long
How long data is kept is set separately for people, vehicles, faces and number plates, and the biometric policies are flagged as such. Sensitive areas are covered with privacy masks, and whoever uncovers one leaves a reason, a time and a name, with a second supervisor’s approval if the organisation requires it.
From a search to evidence
Any result can be attached to the incident and to the case file, with its chain of custody. And what you export is a column file with a fixed header and a dated filename that opens in any spreadsheet: nobody has to trust a screenshot.
What does the real work
Almost all of it is solved without the face
Someone searching a video archive for a person almost never has a photo of their face. They have a description: dressed in red, carrying a rucksack, left by the back door around seven. That is exactly what IRIS notes about every object while it records: what it was, what colour, which way it went, how long it stayed, in which zone and at what time. That is enough to find someone without knowing who they are, and it is what almost every real search comes down to. It is also the least intrusive way to do it, and here it is the main route rather than a consolation prize: biometrics is the exception, and it ships switched off.
A search never rewatches the video: it queries what was written down while the video was being recorded. Every object leaves more than fifty measured details behind, fifteen criteria combine in a single query, and against them you can ask for any of the 94 catalogue classes, any of the eleven colours or a free colour picked on screen, and any of the nine behaviours: fast, stopped, loitering, direction of travel, dwell time, line crossing, wrong way, abandoned object or zone visited. A filter nobody asked for is never filled in for you. The screen always says how many results really exist before the list was trimmed. And when there are none, it says why: no permission, no recording, not indexed, switched off, or genuinely no match. A failure is never dressed up as “there is nothing here”.
Finding the same one on another camera is what turns a search into a route. IRIS proposes candidates ranked by confidence level — high, likely or ambiguous — and publishes the three signals behind each one separately: the likeness, how well the timing fits, and how close the two cameras are. It rules out by itself what would be physically impossible: if there was no time to get from one camera to the other, the candidate is not even shown. And there it stops. The system proposes; a person confirms or rejects, and that decision is stored with a name and a time. Identity is never asserted: the screen says possible match by appearance, and warns that it is no proof of identity and needs human review.
The product includes face recognition. It ships switched off, and it is not offered in the European Union. Where the law allows it, turning it on means going through a screen that cannot be skipped: nine warnings that must be read and three fields that must be typed in — legal basis, purpose and retention period; without all three the process stays inert, and withdrawing the legal basis stops it there and then, with nothing to restart. We do not make that call and it does not come switched on: each site decides, who turned it on and on what basis is on record, and everything anyone searches for is audited — who looked for what, and when. Across the nine zones this site shows, IRIS identifies nobody, because on those cameras this function is not switched on.
An object does not live on a camera: it crosses it. From one sighting you can ask for candidates on neighbouring cameras within the next quarter of an hour, or open the window to 168 hours — a whole week — to follow the same lorry or the same person across the site. Every sighting a person confirms is added to a file for that object: first and last time it was seen, how many cameras saw it, how many sightings make it up and, if it is a vehicle with a reading, its most repeated number plate. That file is the route, and it is also what gets attached to an incident or a case. The fingerprint history reaches roughly fourteen days: the date is not locked, you are simply warned that further back there may be nothing left, because hiding the limit would be worse than saying it.
A system like this is judged by what it does when it finds nothing. Here every empty list gives its reason: nothing reaches the threshold — and it says so with the threshold actually used, not the one now sitting on the slider — the marked object is not indexed yet, there is no permission, there is no recording, or appearance search is switched off for the whole platform, with the exact place where it is switched back on. A row that stands for dozens of sightings says so, with the real interval it covers. And a stationary object does not flood the list: it is measured the first time and then only now and then, because a parked motorbike measuring itself every second used to bury everything else. It stops repeating; it does not stop being indexed.
The figures
Counted, not estimated
- 94object classes you can search for by name
- +50measured details kept for every object that goes past
- 15criteria that combine in a single query
- 9legal warnings that must be read before biometrics can be switched on
- 11catalogue colours, plus any free colour picked on screen
- 168hours of window to follow the same object across cameras: a whole week
Questions
Does IRIS recognise faces?
The product includes face recognition, it ships switched off, and it is not offered in the European Union. Where the law allows it, turning it on means reading nine legal warnings and declaring a legal basis, a purpose and a retention period: without those three it does not run, and withdrawing them stops it there and then, with nothing to restart. Each site decides; it never arrives switched on, and who switched it on and when is on record. In the zones this site shows, IRIS identifies nobody, because on those cameras the function is not switched on. Almost every real search does without it: type, colour, route, dwell time, zone and hour are enough.
Can a person be found without knowing who they are?
Yes, and it is the normal way to work. IRIS keeps a record of what each object looked like — type, colour, size, route, speed, dwell time and zone — and that is enough to find it again on another camera on another day without ever naming it. When it proposes that two sightings are the same object, it says so as a confidence level, not as an identity: the screen warns that this is a possible match by appearance, that it is no proof of identity and that it needs human review. A person confirms or rejects each proposal, and that decision is stored with a name and a time.
Who is allowed to search, and what is recorded?
Searching requires permission to view recordings, and on top of that each user only finds objects on the cameras they may see: a camera without permission does not break the search, it simply contributes nothing, and the screen says so. Every search records who ran it, when, over which cameras and with what criteria. Sensitive areas can be covered with privacy masks, and uncovering one requires a stated reason and, if the organisation demands it, a second supervisor’s approval. How long each kind of data is kept is set separately. And it all happens inside the customer’s own site: the models and the maps are our own and work with no internet, so not one image goes out to a third party. And everything — interface, video, thumbnails and queries — travels over a single port, 443: not one more port to open on the firewall.
What if the person changes clothes?
Then the appearance trail breaks, and that has to be said plainly. A person’s fingerprint rests on colour, shape and build: change the clothes and the fingerprint changes, so IRIS stops proposing that person on later cameras. What it does not do is invent a match: it says nobody reaches the threshold, and states the threshold it used. What still stands is the route up to the change — where the person was last seen and at what time — a search by other criteria in that zone and that time span, and the vehicle if there is one. This is exactly why the route is built from sightings a person confirms, and not from an automatic chain nobody ever looked at.
How long is the appearance fingerprint kept, and who can query it?
Each site sets it, separately for people, vehicles, faces and number plates, anywhere from zero to 3,650 days. Zero does not mean delete: it means keep indefinitely, and the screen flags it in red so nobody picks it by accident. Deletion is carried out by a nightly sweep, and if that sweep is not armed the screen says so instead of assuming the deadline is being met; you can also see how many live fingerprints exist per modality and how many are past their deadline. Querying them requires permission to view recordings, and it only reaches the cameras that user is allowed to see. Changing retention requires its own permission, and every search and every change is recorded with who and when.
The real interface, step by step
You set the rule once. IRIS applies it every time.
You will see a summary of the interface, played step by step and hands-free. Each round starts with another case. The complete tool does not fit in a demo.
- You ask
- IRIS searches
- It shows you
Next step
How long before somebody comes down that corridor at four in the morning?
13 cases: 12 raise an alert when something happens and 1 only measure. You see each image as it is, and then with what IRIS understands on top.
Who sees the child who has fallen at the far end of the playground?
13 cases: 12 raise an alert when something happens and 1 only measure. You see each image as it is, and then with what IRIS understands on top.
And the best part
One camera,and you already have a watchman. Four thousand cameras, and you have four thousand watchmen.
None of them tires, none of them looks away and none of them works shifts. And there is nothing to install: the cameras you already have can start understanding what they see.
Tell us your problem and we will say whether IRIS understands it, or not yet.
We reply the same working day.