Use case
Detecting after-hours access on a site that's supposed to be empty
Once the site closes, anyone inside the line is unexpected by definition. The hard part isn't spotting them — it's telling them apart from the people who are meant to be there.
VuePoint
Detecting a person on a closed site is the easy half. Any competent detector will find them. The reason after-hours systems get switched off is the other half: they find everybody. The night cleaner, the delivery driver at 05:40, the site manager who came back for a laptop, the fox, the security patrol you're paying for. After a fortnight of that, nobody reads the alerts, and the system has made things worse than having nothing.
So the useful question is not "is there a person?" It is "is there a person whose presence nobody can account for?"
Time is the first filter, and it does most of the work
A person walking across the yard at 14:00 is the business operating. The same person, in the same place, at 02:00 is the entire event. Almost all of the value here comes from the schedule being right — which means it has to be the real schedule, not the one on the gate sign.
- Different closed hours for different areas. The yard closes before the office does.
- Weekends and holidays are their own schedule, and they're when most of these events happen.
- Known exceptions are entered in advance — the weekend pour, the overtime shift, the scheduled delivery. An exception that nobody recorded becomes an alert, which is correct.
- Seasonal daylight is irrelevant. Gate the zone on the site's operating hours, not on darkness.
The second filter: where, not just when
Zones matter more than the property boundary. A person on the pavement outside the fence is not an event, however close they walk. A person between the fence and the building is. A person at the material store at 03:00 is a different order of event from a person at the gate, and it should be handled differently.
| Zone | What presence there means | Reasonable response |
|---|---|---|
| Outside the line | Passers-by. Not an event. | Never alert. Excluding this zone properly removes most false alerts on a street-facing site. |
| Approach and gate | Someone is at the entrance. Could be legitimate. | Log it. Alert only on dwell — arriving and leaving is not the same as arriving and staying. |
| Inside the perimeter | Unexpected by definition once closed. | Alert, with the clip attached. |
| High-value areas — material stores, plant, fuel | Highest consequence. | Alert immediately, and keep the sequence rather than a single frame. |
Verification is the part that determines whether anyone trusts it
An alert that says "person detected" makes someone open an app at 3am to find out it was a cat. An alert worth receiving carries enough for the decision to be made from the notification itself: what was detected, where, at what time, how long they stayed, and a clip that shows it.
This matters more than detection accuracy in practice. A system that detects everything and explains nothing gets muted. A system that detects slightly less but tells you what it saw gets used, and a system that gets used is the only kind that helps.
What the record is actually for
Most after-hours detections are not incidents. They are the ordinary business of a site: someone came back for something, a driver arrived early, a patrol did its round. The value of logging all of it consistently shows up later, in three ways that have nothing to do with alerting.
- When something does go missing, you have a bounded list of who was on site in the relevant window instead of four days of footage to search.
- When a contractor's hours are disputed, arrival and departure are recorded rather than remembered.
- When you want to know whether the patrol you're paying for is happening, the record answers it without anybody having to ask.
On sites that change layout constantly, the zones and schedule have to move with the work — see what actually breaks on temporary deployments for the physical side of keeping that accurate, and construction sites for how this fits the rest of what happens on an active job.
Common questions
- Can it tell staff from strangers?
- Not by identity, and we would be cautious about any system that claims to — biometric identification carries specific legal obligations in several states. What it can do is use context: expected hours, expected zones, expected routes, and known exceptions entered in advance. That distinguishes accountable presence from unaccountable presence without identifying anyone.
- What about animals?
- Classification handles most of it, and it is worth checking rather than assuming — the failure mode is usually a cat close to the lens looking the same size as a person at distance. Size relative to the scene at that point, rather than pixel size, is what makes this reliable.
- Does someone watch the cameras all night?
- The analytics run continuously and surface what they find. What happens with an alert is set by your escalation policy — who gets it, at what hours, and what they do with it. That is a decision about your operation, and it should be made deliberately rather than inherited from a default.
Want to know who's on site when nobody should be?
Tell us the site and its real operating hours. We'll show you what the platform would surface and what it would leave alone.