AI cameras tested on UK roads have flagged hundreds of possible seatbelt and smartphone violations. The system photographs vehicles through their windshields, then sends images that may show an offence to a company inspection team for review.
The trials show how computer vision can turn ordinary cameras into detailed monitoring tools. They also illustrate why the process between an automated flag and a report to authorities matters.
How the road camera system works
Aecom and Acusensus tested the system for several months. A camera mounted high above the road takes pictures through the windshield and software checks them for possible traffic violations.
It can identify whether passengers appear to be wearing seatbelts and whether drivers are looking at smartphones. The software can also flag a phone resting on a driver’s lap.
A possible violation does not automatically become a report. The image is sent to an inspection team from the camera company. If that team confirms the violation, it sends a report and the image to the authorities.
This review step makes the system a combination of automated detection and human judgment. The camera narrows attention to images that may show a violation; a team checks them before the information is passed on.
Trials recorded hundreds of possible violations
In an initial trial in Devon and Cornwall last fall, the system detected 630 violations in just over a month, excluding speeding. Of those, 40 drivers were caught using smartphones without permission, while 590 were found not wearing seatbelts.
A separate test in Cornwall found about 300 more violations in just 72 hours. The reported cases included 180 passengers without seatbelts and 117 people using smartphones without permission.
Tests in Hampshire and the Thames Valley in July 2023 found more than 350 violations in one week. The figures included 273 passengers not wearing seatbelts and 86 drivers looking at their smartphones.
These results cover different trials and periods, so they describe what the system flagged in those settings. Taken together, they show the scale of potential violations that a camera system can identify for review in a relatively short time.
Computer vision can track more than traffic
The same general ability to interpret camera images can be used in other settings. The source describes restaurant systems that can count customers and measure how long it takes staff to clean a used table.
Monitoring can extend to individual workers. One example tracks how many coffees café baristas serve during their working hours and how long customers stay. Such information could be of interest to large restaurant chains looking to measure activity and performance.
Computer vision can also support uses that appear more directly beneficial, such as workplace safety. An automated warning system could alert people when a forklift is overloaded.
Across these examples, the underlying capability is similar: software analyzes images to recognize people, objects, or actions, then produces information that someone can use. The effect depends on what is being watched and how the resulting data is used.
Detection brings questions about oversight
The traffic trials focus on seatbelt and phone use, but the broader examples show that AI surveillance can become more detailed. A camera may reveal not only whether an event occurred, but also how long a task took or how often a worker performed it.
The source points to YOLO, an open-source image-analysis system developed by the computer vision community since about 2016. It describes the system as accurate, fast, reliable, and requiring little computing power.
AI researcher Joe Redmon, who helped develop YOLO, stopped working on the software in March 2020 for ethical reasons. He said he could no longer ignore military applications and invasions of privacy.
That concern applies beyond any one camera deployment. As image-analysis systems become capable of recognizing more activity, questions about review, purpose, and privacy become part of the technology’s impact. In the road trials, human inspection is one checkpoint between detection and reporting; other uses may involve different decisions about what to record and how to act on it.