Somewhere between 94% and 99% of the burglar alarms police respond to are false. That is not a rough guess. It is the finding of research published by the U.S. Department of Justice and it has barely moved in the two decades departments have tracked it. The Chicago police log more than 300,000 alarm activations a year, with roughly 98% turning out to be nothing. The DOJ estimates false alarms alone consume as much as $1.5 billion annually in wasted police time, nationwide.
The problem is old enough that the security industry has written formal standards trying to fix it. The Security Industry Association has a technical standard, ANSI/SIA CP-01, built specifically around reducing false alarms in control panels. Video security has inherited the same failure, according to Asaf Dan, principal product manager at AI video surveillance company Lumana.
“They detect motion, not meaning,” he says of legacy and most existing systems. He explains that a falling object triggers the same alert in these systems as a person entering a restricted area. The same for wind, wildlife, headlights and a spider crossing the lens: All of them register as a threat because the system has no way to know otherwise.
Legacy systems trained to be ignored
The cost of that gap is not just wasted alerts. It is trust and once that goes, Dan says, it does not come back easily. “Legacy systems generate false alerts at a rate that slowly trains security teams to stop paying attention,” he notes. While operators burn hours chasing false alerts, genuine incidents get buried in the same noise and the security teams paying for the system stop believing it works.
Part of the problem is structural. Most video AI tools, Dan says, are trained once before deployment and then are frozen, which means they can never learn the difference between a parking lot at noon and the same lot in the middle of the night. These tools have no way to build a picture of what normal actually looks like in the one place it is watching at different times, so they alert on everything equally.
Teaching cameras what normal looks like
Lumana’s answer is a hybrid physical AI platform with the model it calls VIA-1 running on the cloud and the edge, which the company says reduces false alerts by up to 90%. Instead of reacting to any change in the picture, Dan says “the platform can understand what it is actually seeing.” He explains that each camera gets its own model and that model learns the specific rhythm of its own location over time – what a hospital lobby looks like at midday, for example, versus a loading dock in the middle of the night..
He highlights the example of a plastic bag blowing across a parking lot. To a legacy system, this bag and a person walking through the same frame both register as identical movements. Dan says Lumana’s platform reads the shape each one makes and tells them apart so that it flags the person and ignores the bag.
A similar problem shows up with detecting falls: a person bending down to tie a shoe looks like a fall to a legacy system that only tracks motion. Dan explains that the Lumana software waits to see whether the person actually stays down before it creates an alert.
The system also adapts as conditions change. For example, on a windy day, swaying branches would normally set off constant alerts in legacy systems. Dan points to one deployment where a system like this held onto roughly twice the detection accuracy of an older, static model through a full change of seasons, while the static model’s accuracy dropped.
What gets filtered out and what does not
Cutting that much alert volume raises a fair question: Does anything real get missed? Dan says no. He indicates that every event is still recorded in the Lumana platform and is searchable whether or not it triggers an alert. So what changes is what causes an alert to be generated and not what the cameras capture.
He is also clear that eliminating false alerts completely would be the wrong goal. “Zero is the wrong target,” he states. Real environments are complex, so some noise never fully goes away.
Reducing false alerts by 85% to 90% is realistic, in his view, but a system tuned to hit exactly zero risks filtering out real threats. What Lumana is aiming for instead is a level of noise low enough that when an alert does come through, security teams believe it.
What security teams should ask before they buy
For a team already buried in false alerts, Dan’s first suggestion is not to look at new software, but to look at the existing data. Check what share of alerts actually led to actionable activities and find out which cameras are generating most of the noise. In his experience, a small number of cameras usually cause most of the problem.
He also warns against the quick fix of just turning sensitivity down. “That does not solve anything,” he says. Instead, it trades one kind of failure for another: swapping false alerts for missed events elsewhere.
When testing a new AI vendor, he recommends trying it in conditions close to your own, checking that it works on the cameras already installed and asking the vendor to explain, specifically, why any single alert was generated or got filtered out. “If they cannot explain it,” he explains, “that tells you something.”
The number worth tracking after installation is simple: out of all the alerts, how many actually led to action. If that ratio is not climbing, Dan says, the system is doing the same thing every legacy platform already does. It is teaching the people watching it to stop paying attention.
Featured image credit
