A camera estate runs unwatched. That is arithmetic rather than a failure of discipline, because a site with a few hundred cameras rarely has more than a couple of people at the desk. Most of it is never watched, and it is deleted on schedule with whatever it held still inside.
So a case has to start somewhere. Five things can, and they differ in how much you set up beforehand and in how good a starting point each one leaves you. Two of them repay careful configuration and the other three arrive as they are. This sits under our guide to multi-camera triangulation, which covers connecting separate sightings of the same thing into a route.
What a camera raises on its own
The simplest route needs almost no setting up. Each camera is given a list of classes to watch for. Spotting one is itself an alert, carrying the severity you set for that class on that camera. A weapon appearing in frame needs no zone drawn and no rule written. The class alone is the point.
Most places want something narrower, because a person in a corridor is not news and a person in that corridor at two in the morning might be. That is what a rule adds. A rule is a zone drawn on the camera's own frame with a condition attached to it, and four kinds ship.
| Rule |
Fires when |
What you set |
| Loitering |
Something stays inside the zone too long |
The dwell time |
| Intrusion or line crossing |
Something enters the region, or crosses the line the way you drew it |
The zone shape, or the line and its direction |
| Crowd count |
Too many of a class are inside the zone at once |
The capacity |
| Direction of travel |
Something moves against the expected flow |
The expected direction |
Two details decide whether a rule works or floods you. The first is that each rule names the class it applies to. One camera can therefore run a loitering rule on people and an intrusion rule on vehicles at once, each with its own zone. The second is that time inside a zone is treated as one continuous stay rather than a run of separate entries. A person standing in a doorway for ten minutes produces one alert instead of forty.
One limit governs all four. A rule is scoped to a single camera and a zone does not span cameras. A place with two approaches needs a rule drawn on each of them. That limit is not an oversight, and the reason for it explains the next route entirely.
Conditions that span cameras
A zone rule cannot cross a camera boundary because of where it runs. It is evaluated inside the live pipeline, one detection at a time, as frames arrive. At that moment the only thing available to compare against is the frame in hand.
A condition works the other way round. It runs over events that have already been written to storage, so every camera's events are available to it at once. That is what lets it cross both camera boundaries and time windows, and it is the only reason the three shapes below are possible.
The first shape matches one thing together with another, such as a specific license plate detected alongside a person carrying a specific attribute. The second requires several cameras each to show a situation, so cameras A, B and C each meet a condition you define. The third waits for a situation to persist, such as two cameras showing given situations continuously for ten minutes.
The delay is what you pay for that reach. Because evaluation follows recording rather than happening inside it, a condition sits above the real-time alerts rather than replacing them. Anything that has to be acted on in the moment belongs in a zone rule. Anything that could never be seen from a single frame belongs here.
Tuning is where these go wrong, and they go wrong in both directions. A condition written loosely fires often enough that whoever receives it stops reading, which is the same as having no condition at all. Written tightly it may never fire, and no warning tells you so. Months pass with a rule that was never going to match. Write the loose version first, watch what it produces for a week, then tighten it against real traffic rather than against a guess. A match notifies by email, or by a web request into a system you already run. The second can drive a process rather than a person, which is what turns a condition into something other than one more thing to read.
When a person tells you instead
The remaining three routes need no setting up, and the loosest of them is a colleague walking in to tell you. A rule knows the second it fired. A witness gives you this morning, or around three, or sometime last week, and usually says so hours or days after the fact. The window you search is therefore a guess, and it has to be wide enough to hold the truth and narrow enough to be worth searching.
Build that window rather than picking one. Start from what the person actually said instead of a round number. Then widen it by however far their memory is likely to be out, which for an afternoon reported the next day is easily an hour either way. Then add a margin for the clocks. If your recorders and whatever the person read the time from do not agree, the window is aimed slightly wrong before anyone opens a recording.
A window that returns too many people is not always fixed by narrowing it. Some places are busy at some hours, and no version of this produces a short list there. The same place at a quiet hour usually does.
Both a report and a rule end at the same question, which is which cameras were looking at that place. A rule answers it, because it fired on a particular camera. A report does not, so the security team has to know that the loading bay is covered by three specific cameras. That belongs written down once rather than reconstructed under pressure by whoever is on shift. Some places have no camera and never will, because they are places no one should be watched in, and a room that cannot be covered still has doors.
When a name is already on a list
The last route is the only one that arrives with a name attached. Where a person has been enrolled with their photograph, each camera running detection compares the faces it sees against that gallery. A match raises an alert. The same shape applies to a vehicle whose plate is on a watchlist.
No zone has to be drawn and no behavior has to be unusual. They appear, and you are told. It is also the only route needing a decision made in advance about who belongs on that list, and on whose authority. Finding every recording a person appears in covers enrolment in full.
What you have once something reaches you
Four of the five routes leave you in the same position. A time, a place, and no person. That is a starting point rather than an answer, and cross-camera tracking covers following whoever was there across your other cameras and back through your recordings. The gallery match is the exception, because it hands you the name with the alert.
Whatever the route, what reaches you is a condition in a place rather than a statement about a person. A rule reports that something happened where you were watching. It does not report what anyone did there, and that distinction belongs in writing before anyone is asked to act on one.
What not to watch, and who decides
Rules are switched on per area, which makes this a live decision rather than a paragraph in a policy document that is never opened. Decide which spaces stay uncovered and write the answer down. Decide who may configure a rule. A zone drawn on the wrong doorway is a decision about surveillance taken by whoever happened to hold the login that afternoon.
Settle the harder cases before anything is configured rather than after the first objection. Two come up everywhere. A space deliberately left unwatched stops being unwatched the moment a rule is drawn on the camera covering its doorway. And footage that records a person can usually be demanded by that person or by someone acting for them. What you keep is therefore answerable to somebody outside your organization. Neither of those depends on the sector you work in, and both are cheaper to answer in writing now than under pressure later.
A record that holds up later
If any of this informs a decision about a person, the record has to survive being questioned months afterwards. For a rule that means what fired, on which camera, at what time, and what the rule was set to. For a condition it means the condition as written, because the same match means different things depending on how loosely it was drawn. For a report it means what was reported and by whom, the window that was searched, and why it was that wide.
Then in every case: which recordings were reviewed, by whom, and when, and what was concluded, stated as presence rather than as attribution. A record in that shape can be handed to a parent, a district office, an insurer or counsel. A screenshot with a name written under it cannot.
Where to begin
Take last month's reports and ask, for each one, whether a rule would have caught it. The answers tell you which rules are worth drawing and which places are worth covering, which is a better basis for that decision than a list of features. Two checks cost nothing and go alongside it. Confirm your recorders agree on the time, because every window you draw is aimed by it. And put the camera list on paper before the security team needs it at speed. Rules run on the cameras and video management system already in place, through VIDIZMO AI Live Insight, so none of this starts with new hardware. For what a camera can be set to detect in the first place, see AI video analytics. For how this sits in a wider camera program, see video intelligence.
