By Sam P. · 8 November 2026 · 4 min read
How AI dashcams detect phone use and no seatbelt
An AI dashcam detects phone use and a missing seatbelt with a second camera pointed at the driver and a small computer-vision program running on the camera itself. The program looks for shapes it has been trained to recognise, such as a hand holding a phone or the absence of a belt strap. If it sees one for long enough, it saves a short clip.
One point before the detail. The cameras are third-party units. AutoM8 integrates and supports them, but we do not make the camera hardware.
Spotting a phone
The inward-facing camera looks for a hand near the face holding something phone-shaped. The program has been shown huge numbers of labelled driving clips, “phone” and “no phone”, until it can pick out the pattern unaided. It works a bit like a spell-checker, flagging whatever does not fit.
All of this happens on the camera, not in a distant data centre. That is called edge processing, and it means the camera can react within a second or two without a good mobile signal.
To avoid false alarms, most cameras only flag phone use when it lasts a few seconds and the van is above a set speed. Glancing at a screen in a car park is a different thing from scrolling at 50mph, and the thresholds are tuned to separate the two.
Spotting a missing belt
The same camera checks whether the diagonal strap is visible across the driver’s chest. No strap while the vehicle is moving raises a no-seatbelt event. Many units add an infrared sensor, which is light your eyes cannot see, so the strap and the driver’s posture stay visible in a dark cab or at night.
Again the event waits a few seconds of movement. That stops the normal gap between starting the engine and clipping in from counting as a breach.
What makes the camera save a clip
The camera records constantly but keeps very little. It holds a rolling buffer of the last few seconds and only commits a clip for upload when something crosses its confidence threshold: phone use, no belt, signs of distraction, or a jolt that looks like a collision.
That keeps review short. Rather than hours of nothing, you get the handful of moments that matter, usually available the same day.
Linking a clip to the journey
Every saved event carries a timestamp and a location, so it appears on the same live map as the rest of the vehicle’s journey in AutoM8. You see the no-seatbelt event, then exactly where on the route it happened and what the van was doing.
The events also sit beside driver behaviour scoring, which judges harsh braking and speed against the vehicle’s weight, so a video flag and a score change point at the same trip. The camera side is covered on the AI dashcams page.
How often they get it wrong
They are good, not perfect, which is why the design leans on time and speed thresholds and why a person reviews flagged clips. A “phone” can turn out to be a sandwich. The camera surfaces moments worth a look and does not hand down a verdict.
For a fleet, that changes the conversation with drivers. You point at a specific clip from a specific journey rather than making a vague complaint, which makes coaching fairer and easier to act on. To watch it run on your own vehicles, book a demo.
Related features