A person came to attention on one of your cameras today, and the next question is whether they have been here before. That same camera was running in March, and so were the others. Nobody was watching them at the time, because a site with a few hundred cameras records far more than anyone can watch. What those cameras left behind is footage nobody has opened. So you go looking for the week around it.
The search returns nothing. An empty result and an absence look identical from where you are sitting, and they are not the same thing at all. One means the person was not there in March. The other means the analysis that ran over that week was never able to tell you either way. This sits under our guide to multi-camera triangulation, which covers connecting separate sightings of the same thing into a route.
What analysis already did to your footage
A recording is not searched directly, because nothing in a video file responds to a query. A video file is a sequence of images and a soundtrack, and neither of those is text. What a search reads is the output of the analysis that ran over that footage, which is stored alongside it. Where no analysis ran, a query has nothing to touch.
Spoken audio is what usually gives a recording something to search. A meeting, a call or an interview produces a transcript, and a transcript is text a search can match a name or a phrase against. Surveillance footage carries no speech worth transcribing. A camera over a loading bay records a soundtrack of engines and wind, so the route that makes most video searchable is closed before you begin.
Three kinds of analysis stand in for that missing transcript, and all three ran on the footage rather than on the camera. Detection names what was present, each thing with a class and a moment against it. On-screen text is read and indexed, which is how a license plate or a sign becomes searchable as words. Visual description is the third. A model looks at the frames where the picture changes and writes what it sees, in language, timestamped across the recording. Those descriptions are indexed alongside the transcripts and the read text.
So the search depends on what the analysis found, not on what the camera actually saw. Something was almost certainly found, because detection was running and its output was kept. Whether it covered the thing you now care about is a different matter. That was settled by which classes were being detected, and by where the thresholds for them sat.
What governed the analysis that ran
Detection covers a broad set of classes. It runs to people, vehicles, weapons, protective equipment, devices and screens, identity documents, traffic signs and signals, and environmental hazards. Each camera names the classes it watches for, so different cameras watch for different things, and how an incident on camera becomes known covers where those are set. That list is the ceiling on what the analysis could have found in anything a camera recorded. A camera over a loading bay set to watch people and vehicles found nothing about the bag left by the door. Nothing in its output mentions the bag, however carefully you search.
Within the classes a camera is watching, a confidence threshold decides which detections count. It is set per class on each camera, so one camera can watch vehicles at one sensitivity while another watches people at a different one. Raise it and fewer detections survive, including some that were real. A figure glimpsed at the edge of a frame in poor light stops appearing at all. Lower it and more get through, including things that were never there. Precision and recall are a tuning decision rather than a fixed property of the system.
A second setting works differently and costs something different. A detection has to persist across a number of consecutive frames before it is kept, and that number is configurable too. Raising it suppresses false positives that appear in a single frame. It pays for that with a short delay before a genuine detection is confirmed, rather than with the detection itself. So the two are not interchangeable. One decides whether a marginal thing is picked up at all, and the other decides how quickly a real one is confirmed.
None of these was set with your case in mind, because none of them could have been. What the analysis found therefore varies from one camera to the next and from one class to the next. An empty result is not evidence that nothing happened. It is evidence that nothing was found, which is a weaker statement and an easy one to mistake for the first.
Running the analysis again
What ran in March is not all that can ever run. A recording can be analyzed again afterwards, and whatever that finds joins what is already there. Analysis runs in Nexus, so a period has to be held there before anything can be run over it. Footage that is only in your video management system is out of reach. Once a period is in place, analysis is requested per recording and per capability. So a camera that was not detecting bags in March can still yield a record of them now, from the recording it made at the time.
That only helps with a class you now know you want. Running it again for bags finds bags, and it finds nothing about the thing nobody has thought of yet. So the value of going back depends on how specific your question already is. A case that has narrowed to one object gains a great deal from it, and a case still asking what happened gains very little.
Deleted footage is the one gap that nothing fixes. Retention decides when that happens, and it applies in both of the places footage can sit. Where the window is shorter than the time a case takes to surface, the week you wanted was gone before anyone thought to ask for it. Once the analysis has run on whatever survives, its output holds detections for every class you asked for, and a search of that week's recordings will return them. What comes back is a proposal rather than a finding, because detection picks up things that were never there and misses things that were.
Correcting what the analysis got wrong
Detections ride alongside a recording as a timed track rather than being painted into the picture. Each segment of video is paired with a file of cues naming the classes seen in each moment and how many of each. Turn on the overlay and they appear as boxes over the frame, and on the timeline under the player they appear as markers. A week of recording shows where the activity was without anyone scrubbing through it. The video itself stays clean, which matters when the footage leaves the building. What a prosecutor receives should be the recording, not the recording with a model's guesses painted across it.
Every detection it produced can be changed by a person. A bounding box in the wrong place is adjusted, and one thing recorded as two is merged into one. A track that ran two things together is split apart. Objects are renamed, false positives are deleted, and things the detector missed are added by whoever saw them.
A recording held in the library carries the library's access control, its retention policy and its chain of custody. That is what lets footage be handled as evidence rather than as camera output. Correction sits on top of it. Automated detection proposes and a person disposes, and correcting it turns a machine's output into a record somebody is accountable for. Nobody reviewing the case months later has to take the detector's word for anything, because a person went through it and said so.
You hold a corrected account of what the cameras had in view. It covers the days you chose to analyze and the classes that were being detected, and it covers nothing else. Neither limit is visible in the result itself, which looks the same whether it came from one camera or from forty. Both of them travel with it into whatever decision gets made next, so both belong written down beside it.
What a result is worth
What comes back, once it has been corrected, is worth having. It says this person was in the loading bay at two in the afternoon on a day in March. Behind that sit a camera, a timestamp and a person who reviewed the detection and let it stand. Nobody's memory is in it and nobody reconstructed it afterwards. The recording carries the library's access control and its chain of custody, so it arrives as evidence rather than as camera output.
A negative result is the one that has to be handled carefully. Finding nothing in a week of recordings is not the same as nothing having happened that week. It may mean nothing was detected, or that nothing cleared a threshold, or that no analysis ever ran over that week. Written into a report as "we searched the week and found nothing," it reads as a finding when it is not one. That is the sentence a case comes apart on.
So a result is worth stating with its boundaries attached: which periods were analyzed, which classes were being detected, and where the threshold sat. Those three turn it into something a reviewer months later can rely on, and they cost a line to write down. What no analysis will tell you, at any boundary, is why anyone was there. Our guide to multi-camera triangulation covers what presence does and does not establish.
For how one appearance is connected to another across your cameras, see cross-camera tracking. Where you already have a name and a photograph, finding every recording a person appears in covers what enrolment changes. For how this sits in a wider camera program, see video intelligence.
