Smart Traffic Cameras: How They Detect Phones and Seatbelts

You are on your way to work when your phone buzzes. You lift it to chest height for two seconds to read the message. A few days later a fine arrives, and your first question is not "did I do it?" It is: who actually saw me, a person or a machine? And how far should I trust it?

I have worked on camera-based traffic monitoring systems: capturing the image, reading plates, and designing the review steps that sit between a camera and a fine. Here I explain in plain words how these systems work, keeping what I know from experience separate from what is publicly documented. I am not describing any one vendor, because details differ. Everything is current to October 2026, and every number has a source at the end.

From camera to fine: what happens in a fraction of a second

The process has four stages, and each one discards what does not need to go further:

1. Capture. Several cameras and an infrared flash capture clear images of passing vehicles in all conditions. That is how New South Wales in Australia describes its system, one of the best-documented public programs. Queensland says its cameras take several images of every passing vehicle, covering the plate and the two front seats.
2. Detection. Software finds the windscreen in the picture, then the driver, then the hand and the chest.
3. Classification. The model is asked: is that a phone or something that looks like one? Is the belt running diagonally across the chest? Its answer is not yes or no but a confidence score, such as 92% or 40%.
4. Review and decision. Anything above a chosen confidence level goes to human reviewers; the rest is deleted.

Stages one and four are publicly documented; two and three are my general description from experience, and the details differ from system to system.

That confidence threshold is the dial most people never think about. Lower it and more images reach reviewers while fewer violations slip through. Raise it and reviewers have less to do, but more real violations are missed. There is no perfect setting; each authority picks its balance between the two kinds of error.

Scale explains why the machine is unavoidable. According to an official Transport for NSW review, its mobile phone detection cameras carried out about 260 million vehicle checks between March 2020 and the end of 2022, and over 99% of those vehicles were compliant. No human team can look at 260 million photos. The AI's job is not to catch you; it is to say "not this one" quickly to millions of images, so only a few ever reach a person.

How a model "sees" the phone in your hand and the belt on your chest

The model does not understand that you are reading a message. It learned from large numbers of images that people had labelled in advance ("a phone in a driver's hand", "a fastened belt") that these situations have recurring shapes, and it estimates how closely a new image resembles them.

For a phone it looks for a small rectangle near the ear, in the hand or in the lap; in my experience a head tilted downward is another common cue. For a belt it looks for a diagonal band from shoulder to hip across the chest, and tries to tell it apart from clothing colours and shadows. That is both its strength and its weakness: it is good at clear, repeated patterns and stumbles when the view is blocked. A rectangle in a hand might be a phone, or a wallet, or a glasses case.

Where do Jordan and Saudi Arabia stand?

Jordan. In the statements I could find, the body speaking for the capital's smart camera project is the Greater Amman Municipality. Its executive director of traffic said in November 2025 that the project covers about 5,500 cameras, 2,500 of them in the first phase, and that violation cameras make up no more than 30%; of those, between 1,500 and 1,600 are dedicated to speed, red-light and average-speed enforcement. Roya News reported on 12 February 2026 that new cameras equipped to detect phone use, missing seatbelts and speeding had been installed at eight sites, including the Fourth Circle, Salt Road and Dabouq.

Saudi Arabia. Automated enforcement of seatbelts and phone use officially began in March 2018 in Riyadh, Jeddah and Dammam, after two months of warning messages only. A peer-reviewed study published in 2018 describes how the traffic department trained and employed 120 people at its citation headquarters to document and verify violations, and how offenders received a text message with the date and type of violation. The study presents it as a new system that followed Saher, the speed-enforcement system the police adopted in 2010. I found no recent announcement about its current expansion.

Does a human check the violation before it is sent?

It depends on the authority. Here is what is documented:

• New South Wales: a suspected image passes several human stages before a fine: trained vendor staff first; then Transport for NSW officers; then the state revenue agency for the final decision. If the offence cannot be proven, no fine is issued.
• Queensland: the AI filters images, then an authorised officer reviews the image to decide whether an offence took place.
• Saudi Arabia (2018): a team of 120 staff documented and verified violations, as the study reported.
• Amman: I found no description of a human review step, but the executive director of traffic said citizens can object to the validity of a violation and that invalid ones are struck off.

But "is there a human?" is not enough. What matters is whether that person has enough time, a clear standard and the real power to say "no". A review that clicks through thousands of images with one approval button is not a review.

What can go wrong, and why

Errors come in two kinds: a violation that was missed, which worries the operator more than you, and a violation recorded against you when you did nothing wrong, which is the one that matters to you. In my experience three things cause most of it: glare on the windscreen, tinted glass and night.

Tinted glass shows the limits of sight. In the Riyadh study the researchers watched cars by eye and excluded those with tinted windscreens, and they acknowledged that this may have overestimated seatbelt use and underestimated phone use. Those cars stayed outside the numbers.

Some mistakes are documented. In May 2026 ABC in Australia reported two cases. In Perth a driver was fined A$550 because the image showed her shoulder belt under her arm; the penalty was later withdrawn and the department apologised. In Queensland a driver said a passenger's arm covered part of the belt when the photo was taken; her objection to an A$1,200 fine was rejected, and after four court appearances over a year the department dropped the case without explanation. ABC also describes "false positives" where the belt was worn correctly but the image was misread.

In Western Australia the Department of Transport has issued 81,412 seatbelt-related infringements since October 2025, and about 5,700 (roughly 7%) were overturned after review requests, according to ABC in September 2026. Do not read that as the AI's error rate. The report does not break down why fines were cancelled, and part of the appeals concern the rule that makes a driver liable for a passenger's belt. Police there also dropped hundreds of infringements because they could not show the driver knew a passenger's belt was not properly fastened.

The lesson: many "errors" are not in the pixels alone but in the rule used to hold the driver responsible. And to be frank, I found no independent, published error rate for the New South Wales program, or for Amman or Saudi Arabia. Any accuracy figure from a vendor is a company claim unless an independent audit backs it.

Your photos and data: where they go and who sees them

The camera photographs every car that passes, not only offenders, and that deserves a question. Transport for NSW says images the algorithm rejects are permanently deleted and cannot be retrieved, typically within an hour. Images the vendor reviews are cropped and pixelated to remove anything that would identify the vehicle or its location, and those showing no offence are deleted within 72 hours. Queensland says its cameras do not capture or analyse biometric data. In New South Wales an offender can view the image of their offence by entering the notice number.

For Amman, what I found is the report on the February 2026 tender: active storage for 30 days and archiving for a year, with the municipality owning the software, systems and databases. The tender's categories include cameras for detecting missing seatbelts and phone use while driving, and separate cameras dedicated to facial recognition, to comply with privacy and data-protection requirements. I found nothing published on how long violation photos are kept, in Jordan or in Saudi Arabia.

Why these cameras deserve to exist: the distraction numbers

The loudest objection is "these are cameras to collect fines". Whether or not you believe it, start from the danger. The World Health Organization's fact sheet, updated in July 2026, says drivers using mobile phones are about 4 times more likely to be involved in a crash than drivers who are not. A phone slows reaction times, especially braking and responses to signals, and makes it harder to stay in lane. Hands-free phones are not much safer than hand-held ones, and texting raises the risk considerably. A seatbelt, it adds, can cut the risk of death for vehicle occupants by up to 50%.

Do cameras really change behaviour? The evidence says yes, with caveats:

• Riyadh: a study gathered 3,400 field observations before the system (September to October 2017) and about a month after it (April 2018). Seatbelt use rose from 33.9% to 75.8%, and phone use fell from 13.8% to 9.8%. But it is a before-and-after comparison with no control group, made only a month after launch, and the authors call for further studies of whether the effect lasts. Their data also suggest phone use did not fall at inner-city intersections.
• New South Wales: 1.2% of drivers checked were using a phone in the first pilot (January to June 2019); the share who received notices was 0.17% in 2022. The two measures are not identical: one is observed use, the other is notices issued.
• Western Australia: the Road Safety Commission said the rate of seatbelt and phone infractions detected by the cameras fell 76% between September 2025 and August 2026. That figure comes from the commission itself, and ABC notes the cameras were still being trialled in September.

One caveat stays: these are numbers about behaviour, not crashes, and I found no complete, independent evaluation of these cameras' effect on injuries and deaths. The municipality's executive director of traffic said the goal is saving lives, not collecting fines, and the municipality announced that the contractor gets no share of violation revenue. The real test is whether authorities later publish crash figures and error rates with the same transparency.

A practical checklist: before you drive, and if a fine arrives

The bottom line: machines screen, people decide

A camera sees the hand and the phone, but not the reason. Were you reading a message or adjusting the navigation? So the strongest design for these systems is a machine that screens fast, a person who judges against a clear standard, and a driver who knows their right to see the photo and appeal.

That idea, that devices see the event but not the cause, is close to what my book "The Blind Manager" is about: managers who see the reports and the devices but are blind to human dynamics. As for the hand that reaches for the phone at every buzz, chapter 18, "Half Present", of "Resigning from Yourself" takes it up: how screens took away our ability to be here, and how to move the phone out of reach so you can hear the person in front of you.

The next time your phone buzzes while you are behind the wheel, let it buzz.