Nearly every recorder sold in Australia can email you when it sees motion. Almost nobody leaves it switched on for long. Understanding why is the difference between an alerting system you trust and another feature in the menu you stopped using.
What "motion detection" in a recorder actually measures
Classic motion detection compares one video frame with the next and counts how many pixels changed. If the number of changed pixels in a zone crosses a threshold, that is motion, and an event is logged or an email is sent.
The camera has no idea what changed. It cannot tell a person from a shadow, because to the maths they are the same thing: a block of pixels that is now a different brightness.
That is fine in a sealed indoor room at night. Outdoors, it falls apart.
- Rain and hail. Every drop that crosses the lens is a changed pixel. Heavy rain can trigger continuously.
- Insects and spiders. A moth sitting in the infrared glow fills a third of the frame. Spiders build webs across housings and trigger alerts all night, every night.
- Headlights and reflections. A car turning in the street throws a moving pool of light across your wall. Nothing physical entered the site; the pixels say otherwise.
- Trees, grass and shade cloth. Anything that moves in wind moves constantly, and the shadow it casts moves with it.
- Clouds and sun. A cloud crossing the sun changes the brightness of an entire yard in a second.
- Birds, cats, possums, kangaroos. Real objects, real movement, nothing you want an email about at 3am.
Operators respond predictably. First they raise the sensitivity threshold until the obvious nuisances stop, which also means a person walking in daylight may no longer trigger anything. Then they shrink the detection zones until the useful coverage is gone. Then they turn alerts off.
False alarm fatigue is the real failure
A system that sends forty alerts a night and is right twice is worse than no system. Not because the two real events were missed, but because by the time they arrive nobody opens the email. The information was delivered and ignored.
This is why we treat the alert rate as a design number, not an afterthought. If a site produces more than a handful of alerts a night, something is wrong with the configuration, not with the people ignoring it. Our after-hours alerting guide goes through hours, zones, escalation and cooldowns in detail.
What object detection changes
Object detection works differently. Instead of counting changed pixels, a model looks at the image and reports what it recognises and where: a person here, a vehicle there, with a confidence score and a bounding box.
That single change fixes most of the nuisance list, because rain, shadow, headlights and swaying trees are not people. They still change pixels; they are simply not reported, because nothing person-shaped was found.
It also lets you write rules that make sense in plain language:
- A person in the yard between 6pm and 6am, on any day you have marked as closed.
- A vehicle stopped at the gate for more than two minutes.
- A person crossing a line in front of the roller door, in one direction only.
- A count of people through the front door, by hour.
Our own system works this way. A sweep checks the cameras on a schedule, and when a person appears outside the hours you set it emails the people you nominate with the camera frame attached, with a cooldown so one event does not produce twenty messages. Separately, frames are analysed into a searchable timeline of what happened, with short clips captured around each event, so "show me everything near the dock after close on Tuesday" is a search rather than an evening of scrubbing.
Live detection on the dashboard is a different job again: while someone has the live view open, boxes are drawn on the stream the browser is already watching, which adds no extra load on the recorder.
What AI still gets wrong
Being specific about the limits matters more than the marketing.
Pixels still rule. No model can identify a person who occupies twelve pixels. If the camera, lens and lighting do not deliver usable detail at that distance, detection will not either. That is a camera planning problem — see how many cameras you need and the camera count tool.
Partial and odd shapes. Someone crouched behind a pallet, wrapped in a raincoat, or half out of frame may not be reported as a person. Models are trained on common shapes.
Heavy weather and dirt. A lens covered in rain, dust, cobwebs or salt spray degrades detection exactly as it degrades your own ability to see. Cleaning cameras is maintenance, not optional.
Infrared at distance. At night with IR illumination, contrast drops and the usable detection range is shorter than the daytime range. Expect less, or add lighting.
Reflections and screens. A person on a television, in a mirror, or in a window reflection can be detected as a person. Zones usually solve it.
Confidence is not certainty. Detection reports a likelihood. Set the threshold low and you get more nuisance alerts back; set it high and you miss marginal cases. There is a trade-off and it is site specific.
It does not decide anything. The system flags and alerts. A person still decides whether to call someone, attend, or let it go. Nothing here prevents a break-in; it shortens the time between something happening and somebody knowing.
Choosing between them
Classic motion detection still has honest uses: triggering recording in an indoor room, saving storage on a camera that watches a sealed area, or as a crude backup. Our storage guide covers event-based recording, where motion as a recording trigger is reasonable even when motion as an alert trigger is not.
For anything that sends a message to a human — especially outdoors, at night, or on a site with trees, traffic or weather — object detection is the only version most businesses keep switched on past the first week.
If you want to know what your existing cameras could support without replacing them, ask for a quote. We run the same platform on our own warehouse cameras every day. Other common questions are answered on our FAQ page, and the terms used here are in the glossary.