AI video analytics works, inside an envelope the salesperson will not read out to you
AI video analytics for a business camera system is real and it does useful work. The catch is that it works inside a narrow set of conditions, and the serious vendors publish those conditions in documentation nobody reads before signing.
About this piece
- Article
- Security & surveillance
- Written by Nicholas Backwell · Founder, Redsilicon
- Updated 2026-08-24
Everything below the line marked “reliable detection” in that documentation is where the demo happens and where your yard actually lives.
The short version
- Axis publishes hard minimums for its object detection: object size in the frame, light level, and how long a thing must be visible and moving.
- The reliable jobs are simple ones. Counting crossings of a line, flagging how long something stays in an area, and counting how many things are in an area.
- A detector can be right the great majority of the time and still send you an alert stream that is almost entirely wrong, because the thing it is looking for is rare.
- Licence plate reading has published accuracy claims and at least one police audit that came out very differently.
- Weapon detection, behaviour prediction and “90% fewer false alarms” are sold hard and I could find no independent testing of any of them.
The operating envelope, in the vendor's own numbers
Axis Object Analytics runs on the camera itself and is one of the few products where the limits are written down. From the AXIS Object Analytics FAQ, dated June 2024:
| Condition | What the documentation says |
|---|---|
| Human, minimum size | At least 8% of total image height for reliable detection, which is about 86 pixels on a 1080p stream |
| Vehicle, minimum size | At least 6% of total image height, about 64 pixels on 1080p |
| Light | “Our minimum recommended light level is 50 lux” |
| Time visible | Fully visible and moving. Usually 0.5 seconds, at least 2 seconds for more reliable detection |
| PTZ cameras | Analytics suspend while the camera moves, then wait 5 seconds after it settles |
Read those against a real site. Fifty lux is roughly an ordinary corridor. A dim yard at 2am with one wall pack is nowhere near it. A person who has to occupy 8% of the image height will not, at the far end of a lot, which is exactly where you wanted the alert. And if your one clever PTZ camera is patrolling on a tour, it is blind to analytics for most of that tour.
The point is not that Axis is worse than the others. It is that Axis tells you. Plenty of vendors publish no envelope at all, which does not mean they do not have one.
The jobs it does well are boring on purpose
The scenarios the documentation names are crossline counting, time in area, and occupancy in area. Something crossed a line you drew. Something has been standing in a zone longer than you allowed. There are more things in the zone than there should be.
That covers useful work. A gate line that nothing should cross after 7pm. A loading bay where a trailer sitting for four hours means something. A count of people in a retail space.
Notice what none of those do. None says what the thing intends, whether it belongs, or what it is carrying.
Right most of the time and useless anyway
This is the part that decides whether you keep the system switched on, and it has nothing to do with how clever the model is.
Put arbitrary but plausible numbers on a yard. Say the camera sees 500 movement events a night: cats, headlights on the fence, a branch, rain, your own staff. Say a genuine intruder shows up once a month. Now say the detector is good, catches the real intruder, and only misclassifies 1% of the harmless events.
One percent of 500 is five wrong alerts a night. Over a month that is 150 wrong alerts against one right one. The detector was 99% accurate on the thing you measured, and 150 out of 151 alerts in your inbox are still wrong.
That is the base rate problem. No amount of accuracy fixes it on its own, because the rarity of the real event is doing the damage. What fixes it is narrowing what counts as an event: a smaller zone, a time window, a direction of travel, a size filter. Cutting the 500 down is worth more than another percentage point of model accuracy.
An alert stream nobody trusts gets ignored within a fortnight. Then you own an expensive motion detector.
Plate reading, with both sets of numbers
Flock Safety publishes better than 98% plate capture in optimal deployment conditions, and in litigation the company has stated that its cameras accurately capture 93 of every 100 plates. Both figures, and the note that the methodology behind them is not published, are set out by Red Banyan on 5 August 2026.
Against that, Gizmodo reported on 31 July 2026 that an analysis by Roseville police in California of 1,427 alerts sent in 2023 and 2024 found the software had read the plate incorrectly in 71% of them. Flock disputed the framing and pointed at that deployment’s older hardware and non-standard camera placement.
Both can be true at once. High accuracy across every plate the camera ever scans, and a poor hit rate on the small subset that raises an alert, is the pattern the section above describes. The number on the brochure was measured in optimal conditions, and your gate is not optimal conditions.
Claims with nothing behind them
Weapon detection. Behaviour prediction, meaning software that flags someone as suspicious before they do anything. And the flat “90% fewer false alarms” line, which appears in marketing with no test method, no site and no baseline stated. I could find no independent testing of any of them.
If a vendor quotes a number like that, ask what it was measured against, at what site, over how long, and by whom. A vendor with real data will send it.
What this means for your building
A steel building and gravel yard in Bowmanville is where this gets decided. Analytics on the camera at the shop door will work, because the door is lit, people are close and they are walking. The camera on the back fence line will produce noise all night, because the light is below the minimum, people at that distance are too small in the frame, and every passing vehicle on the road behind lights the fence.
The fix is not a better camera. It is a light at the fence, a tighter zone that excludes the road, and a schedule so the rule only runs when the yard should be empty. After that, cloud video with analytics earns its licence fee. Before that, it does not.
What to do about it
- Ask the vendor for their documented minimums: object size in the frame, minimum lux, and what happens on a moving camera. If they cannot produce them, that tells you something.
- Sit with the motion log for a week before turning any rule on, so you know what the camera is really seeing at night.
- Shrink the trigger. Zone, direction, size and schedule, in that order.
- Put light on any area where you expect analytics to work after dark.
- Trial it for a fortnight with alerts going to one person who notes every wrong one. If nobody will do that, do not buy the feature.
- Decide who acts on an alert at 3am. Software that flags a person to nobody is a recording with extra steps, which is the argument for live camera monitoring instead.
Want this scoped for your site?
Tell us the building and what you’re trying to achieve. We’ll tell you what it takes, and whether you actually need it.
Before you call
Is any of this worth it for a 15 person business?
Sometimes. Line crossing on a gate outside hours is useful and cheap. Anything more ambitious usually is not, because the value depends on somebody responding, and small businesses do not have a night shift.
Our current cameras claim AI detection already. Should we use it?
Try it, with the notebook test above. Detection on a camera you already own costs nothing to trial, and the trial tells you more than any spec sheet.
Can analytics replace a monitoring service?
No. Analytics decides what to send. A person decides what to do about it. The first without the second is where most disappointment comes from.