Most procurement conversations about traffic cameras start with the wrong question. People ask how accurate the camera is. The better first question is what it can actually enforce, because automated traffic violation detection covers very different problems, each with its own camera placement, detection method and evidence standard.
A speed camera and a seatbelt camera can both be called “AI traffic cameras.” They share almost nothing beyond the plate reader.
We’ve written separately about how AI traffic enforcement and ALPR work, including what drives read accuracy. This guide goes violation by violation instead. For each one, you’ll see how detection typically works, where the camera goes and what an evidence package needs to contain. One terminology note for readers in the UK and Saudi Arabia: ANPR (automatic number plate recognition) and ALPR (automatic licence plate recognition) mean the same thing. We use ALPR here and explain the difference in the FAQ.
A caveat before we start. What follows describes industry capability. Which violation types a given system detects depends on the deployment, the camera hardware, local law and the models trained for that site. If you’re evaluating any vendor, including us, confirm the exact scope in writing for your project.
What automated traffic violation detection can catch
Across the industry, AI camera systems are used to detect these violations:
- Speeding, at a single point or averaged over a stretch of road
- Red-light running at signalised junctions
- Stop-line and box-junction violations
- Seatbelt non-use by drivers and front passengers
- Handheld mobile phone use while driving
- Wrong-way driving on ramps, one-way streets and divided roads
- Illegal lane use, including bus lanes and restricted lanes
- Illegal parking and stopping in no-stopping zones
- Riding without a helmet on two-wheelers
Some are mature and legally tested in many countries. Others, especially phone and seatbelt detection, are newer and still subject to human review in most programmes.
Speed enforcement cameras
Speed enforcement is the oldest automated violation, and the one with the heaviest regulation around measurement.
There are two broad designs. Spot speed cameras measure a vehicle at one point, typically with radar, lidar or sensors in the road surface, and trigger a capture when the reading exceeds the limit. Average speed (section control) systems use ALPR at two or more points and calculate speed from the time a vehicle takes to travel between them.
The AI part matters less for the speed measurement itself, which usually comes from a calibrated sensor. It matters more for everything around it: reading the plate reliably, classifying the vehicle (a truck may have a lower limit than a car) and matching the same vehicle across two points in average speed systems.
Placement is about sight lines and legal signage. Evidence typically includes the measured speed, the limit in force, a timestamp, location, the plate read and images clear enough to identify the vehicle. Many jurisdictions also require proof that the measuring device was calibrated and approved, so maintenance records become part of the evidence chain.
Red light violation detection
Red light cameras link to the traffic signal controller. According to the Insurance Institute for Highway Safety, the cameras are “connected to the traffic signal and to sensors that monitor traffic flow just before the crosswalk or stop line,” and capture any vehicle that doesn’t stop during the red phase. IIHS notes many programmes allow a grace period of up to half a second after the light turns red.
The stakes are real. IIHS counts 1,119 deaths in red-light-running crashes in the US in 2024, and about half of those killed were pedestrians, cyclists and people in other vehicles.
Traditional systems relied on loops in the road. AI video analysis adds the ability to track each vehicle’s path through the junction, which helps separate a genuine violation from a vehicle that was already committed when the signal changed, or one turning where turns are permitted.
A good red-light evidence package usually shows the vehicle before the stop line with the signal red, and again in the junction, plus signal phase data and time since the light changed. Short video clips are increasingly common because they answer disputes faster than stills.
Seatbelt detection cameras
Can AI cameras detect seatbelt violations? Yes, and several authorities now enforce it. The camera looks down into the cabin through the windscreen, usually from an overhead gantry or a pole-mounted unit, and a model looks for the belt across the driver’s or passenger’s torso.
This is harder than it sounds. Glare, tinted glass, dark clothing and night-time conditions all get in the way. Infrared illumination helps. So does multi-angle capture.
New South Wales in Australia is a useful public example. Transport for NSW says its cameras use multiple cameras and an infra-red flash, and AI software “automatically reviews images and detects potential offending drivers.” Seatbelt enforcement through those cameras began on 1 July 2024. Worth noting for procurement teams: the AI doesn’t issue fines on its own. Images go through several stages of human review before a penalty is issued.
That human-in-the-loop design is, in our view, the right default for any in-cabin violation.
Mobile phone use detection
Mobile phone use detection works on the same in-cabin imagery as seatbelt detection, and often on the same camera. The model looks for a phone in the driver’s hand or held to the head.
False positives are the main risk. A driver scratching their ear, holding a wallet or eating can look like a phone call in a single frame. That’s why serious programmes pair the AI with manual verification and strict rules about what counts as an offence under local law.
Privacy handling matters here too, because these cameras capture images of people inside their cars. The NSW programme says images with no evidence of an offence are deleted, typically within an hour, and that reviewed images are cropped and pixelated to remove details that would identify the vehicle or its location. Plate-reading systems work differently. According to UK police guidance, ANPR stores a record for every vehicle passing a camera and keeps it for one year. Whatever system you deploy, define retention rules for non-violation images before go-live, not after the first complaint.
Wrong-way driving and other lane violations
Wrong-way detection uses video analytics to track a vehicle’s direction of travel against the expected flow of a lane or ramp. It’s less about issuing fines and more about alerting operators quickly, because a wrong-way driver on a highway is an emergency.
That makes latency the key requirement. An alert that arrives a minute late is almost useless. Wrong-way systems are typically placed on off-ramps and at the entry to one-way sections, where a driver’s mistake can still be intercepted.
The same trajectory-tracking approach extends to other violations:
- Bus lane and restricted lane use, based on vehicle class, lane position and time of day
- Illegal U-turns and prohibited turns at junctions
- Stop-line and box-junction violations, where a vehicle stops inside a marked area
- Illegal parking and stopping, based on how long a vehicle stays inside a defined zone
Helmet detection on two-wheelers is also used in some markets with heavy motorcycle traffic. It follows the same logic as seatbelt detection: identify the rider, then check for the required safety equipment.
What makes the evidence hold up
Detection is only half the job. An alert that can’t be defended in an appeal is noise. Across violation types, the evidence packages that hold up have a few things in common: clear images of the vehicle and plate, accurate timestamps, location, the rule broken, and a record of who reviewed it and when.
Why edge processing matters for roadside units
Many roadside cameras run their models on a processor inside or next to the camera instead of streaming everything to a data centre. That cuts bandwidth, keeps detection running during network drops and gets alerts out faster. We explain the trade-offs in our guide to offline-first and edge software.
The models themselves need training on local conditions: local plate formats, vehicle mix, lighting and weather. Our AI as a Service team builds and tunes computer-vision models for this kind of site-specific work, on Azure, AWS or on-premise infrastructure.
Audit trails and human review
Every detection should leave a trail: what the camera saw, what the model decided, what a human reviewer changed and what was finally issued. The Atomquark Traffic Management System combines ALPR cameras with AI violation detection, automatic capture and classification, a live dashboard with sub-second alerts, audit trails and automated reports. It also supports natural-language analytics queries, remote monitoring and over-the-air updates, so software updates reach roadside units without a site visit.
What we run in production at Hazen
At Hazen in Saudi Arabia, our system runs on 12 cameras and handles more than 14,000 detections a day, with 99.5%+ uptime since January 2026. Uptime is the unglamorous number that matters most in enforcement: a camera that’s offline isn’t enforcing anything, and gaps in coverage weaken the whole programme.
Frequently asked questions
ANPR vs ALPR: what’s the difference?
There’s no technical difference. ANPR (automatic number plate recognition) is the usual term in the UK and much of Europe, while ALPR (automatic licence or license plate recognition) is more common in North America. Both describe cameras and software that read vehicle registration plates automatically and check them against records, such as watchlists or permit databases.
Can AI cameras detect seatbelt violations?
Yes. Cameras mounted above the road capture images through the windscreen, often with infrared flash, and a model checks whether the belt crosses the driver’s or passenger’s body. New South Wales began enforcing seatbelt offences with these cameras on 1 July 2024. In the NSW programme, flagged images go through human review before any fine is issued.
How do red light cameras work?
Red light cameras connect to the traffic signal controller and to sensors that monitor traffic near the stop line. When a vehicle crosses the line during the red phase, the system captures images or video showing the vehicle, the signal state and time since the light changed. Many programmes allow a short grace period, up to half a second, before capturing a violation.
How accurate is automated traffic violation detection?
It varies by violation type and site conditions. Speed and red-light detection are mature and rely on calibrated sensors and signal data. In-cabin detection of seatbelts and phones is harder, because of glare, tint and ambiguous hand positions, so most authorities keep human review in the loop. Ask any vendor for accuracy figures measured on your own roads, not lab benchmarks.
Which violations does the Atomquark Traffic Management System detect?
The system combines ALPR cameras with AI violation detection, automatic capture and classification, live alerts and audit trails. The specific violation types depend on each deployment, camera hardware and local rules, so we confirm scope project by project. If you need a violation type that isn’t in standard scope, our AI team can train and validate a model for it.
Do AI traffic cameras store images of every car?
It depends on the system and local law. Plate-reading systems often store a record of every passing vehicle; UK police ANPR data, for example, is kept for one year according to UK police guidance. In-cabin programmes like the one in New South Wales delete non-offending images quickly. Set retention rules for non-violation data before deployment.
Scoping an enforcement programme
If you’re scoping an enforcement programme and need to know exactly which violations a system can cover on your roads, bring us your junctions, corridors and violation list. Talk to Sales and we’ll tell you what’s in scope today and what would need a custom model.