What computer vision can and can't see on a truck, which use cases are worth paying for, and what I learned helping a large trucking operator put camera-based inspection into daily use.
Avanti Technologies · For fleet, maintenance and safety leaders · 11 November 2026 · 10 min read
Automated truck inspection uses fixed cameras, often with lighting, laser and thermal sensors, to photograph a vehicle as it drives through a portal. Computer vision models then flag tyre damage, leaks, broken lights and body damage, and read plates and trailer numbers. It works well for finding visible defects quickly and for settling damage disputes. It does not replace the driver's legal pre-trip check or a mechanic's brake inspection, and most of the effort goes into the defect definitions, site placement and maintenance integration rather than the model.
A while back I worked as solution consultant on a vehicle inspection programme for one of the bigger trucking companies. I can't name them or share their numbers, because the work is under NDA. What I can share is how the thing actually went: what we got wrong at the start, what turned out to matter, and where the technology is genuinely good versus where the sales deck is ahead of reality.
I've also gone back through the public research, because a lot has moved since. Drive-through scanners that started in car dealerships and rental returns are now being built for Class 8 tractors and trailers, and yard operators are putting cameras on every gate lane. So this is part field notes, part map of the market as it looks in late 2026.
The case starts with roadside data. During the CVSA's 2025 International Roadcheck, inspectors across North America ran 56,178 inspections in three days and put 18.1% of vehicles out of service. Brakes and tyres made up most of the vehicle violations. These are mostly things you can see, or at least screen for, before a truck leaves the yard.
The other driver is the paper inspection itself. In the US, drivers file a vehicle inspection report under 49 CFR 396.11; in the UK and Ireland it's the daily walkaround check. Both depend on a tired person with a torch doing the same checklist every morning. Most do it well. Some tick boxes. Nobody can see under a trailer properly while standing next to it.
Then there's damage. Trailers move between drivers, depots, customers and hauliers all day. When a dent turns up, the question is who did it, and without a time-stamped record from each handover, the answer is usually whoever can't prove otherwise.
Strip away the branding and most systems look the same. A truck rolls through a portal or gate lane at low speed. Induction loops or a lidar trigger fire the cameras. Side and top cameras, plus a line-scan camera in a ramp or trench for the underbody, capture thousands of images with strobe lighting so the result doesn't depend on the weather or the time of day. Some setups add a laser line for tread depth and a thermal camera for wheel-end temperature.
The images are stitched per unit and matched to an identity: plate, trailer number, fleet number. Detection and segmentation models then look for specific defect classes. The good systems also compare against the previous scan of the same vehicle, which is how you tell new damage from old. The output is a condition report tied to the vehicle and a time, and ideally a work order in the maintenance system.
The detection is done by trained vision models, not chatbots. Vision-language models are useful at the edges: writing the work-order text from a set of flags, letting a supervisor ask "show me every trailer with a sidewall flag this week", or helping triage the odd image the classifier isn't sure about. Using a large general model as the main detector would be slower, more expensive and harder to audit. It's the same point we made in the most successful AI systems use the least AI.
This is my read, combining what we saw in our deployment with what vendors publish. Vendors quote detection rates in the mid-90s; treat those as their numbers until you've measured on your own trucks.
| Use case | What the cameras look for | How ready it is |
|---|---|---|
| Tyres and wheel ends | Sidewall cuts and bulges, flats, mismatched duals, missing or loose wheel nuts, tread wear | Good for visible damage. Tread depth from images still needs checking against a gauge. |
| Underbody | Fresh leaks, hanging air lines, damaged suspension parts, debris | Strong, and the one drivers can't do well in a car park at 5am. |
| Brakes (thermal) | Cold wheel ends that suggest a brake isn't working | A screening signal. Good at telling you which truck to look at, not a brake test. |
| Lights and reflectors | Dead lamps, cracked lenses, missing conspicuity tape | Easy wins, if the lights are switched on as the truck passes. |
| Body and trailer damage | Dents, tears, panel damage, compared against the last scan of the same unit | Mature. Change detection is what makes it useful for disputes. |
| Identity at the gate | Plates, trailer and container numbers, USDOT and MC numbers, hazmat placards | Mature OCR. Dirty plates and hand-painted numbers are the usual failures. |
| Cargo securement | Strap count and position on flatbeds, tarp condition | Early. Useful as a prompt for a human, not a verdict. |
Two of these deserve a note. Thermal brake screening has the best public evidence of anything here, and it's old. An FMCSA field evaluation from around 2000 found that when inspectors used an infrared camera to pick which trucks to pull in, the share placed out of service for brakes rose from 34% to 84%. It worked at walking pace and failed at highway speed. That tells you exactly how to use it in a yard: as a screen that sends the right truck to the workshop.
Tread depth is the opposite. There's plenty of marketing and very little independent testing of depth measured from images. The gauge in the groove is still the legal reference, so use the cameras to spot uneven wear and flag tyres for a gauge check rather than to sign them off.
The market splits roughly three ways. Drive-through condition scanners such as UVeye came from dealerships and car rental and are now extending to Class 6 to 8 trucks and buses; Amazon has used UVeye across its delivery partner network. ProovStation has the same idea in Europe, mostly in remarketing and light commercial vehicles. Yard and gate platforms such as Outpost, Kaleris and Terminal Industries start from identity, reading plates, trailer numbers and USDOT numbers at the gate, and are adding damage capture on top. Then there are general vision platforms and integrators who will build a custom pipeline if your problem is unusual.
For a large fleet, the build-or-buy question is mostly about the camera hardware and site work, which I'd nearly always buy, versus the defect logic and integration, which is where your own operating rules live and where a partner or in-house team earns its keep.
None of these are about model architecture. The models were the easy part.
One more for European fleets: cameras that see into the cab capture faces, which brings GDPR into scope. Blur at capture, keep retention short, and write it down in a data protection impact assessment before the first site goes live. Our GDPR guide covers the basics.
The metrics that convinced our client's leadership weren't model accuracy. They were roadside out-of-service results, how long a known defect stayed open, how many damage claims got written off for lack of evidence, and how many minutes the morning check took.
It is a set of fixed cameras, usually with strobe lighting and sometimes laser and thermal sensors, that photograph a truck or trailer as it drives slowly through a portal or gate lane. Computer vision models then flag defects such as tyre damage, fluid leaks, broken lights and body damage, read the plate and trailer number, and produce a time-stamped condition report linked to that vehicle.
No. In the US the driver vehicle inspection report under 49 CFR 396.11, and in the UK and Ireland the daily walkaround check, remain the driver's and operator's responsibility. Camera systems make those checks better by catching things a person struggles to see, especially under the vehicle, but they support the legal inspection rather than replace it. Brake adjustment still needs a mechanic.
The most reliable are visible ones: sidewall cuts and bulges, flat or mismatched tyres, missing wheel nuts, fresh leaks and hanging air lines under the vehicle, dead lights, and dents or tears on the body and trailer. Thermal cameras can screen for brakes that are not working by spotting cold wheel ends. Tread depth from images and cargo securement are less mature and should be confirmed by a person.
Vendors commonly quote detection rates in the mid-90s percent, but there is little independent testing, especially for tread depth. The strongest public evidence is for thermal brake screening: an FMCSA field evaluation found that using an infrared camera to choose which trucks to inspect raised the share put out of service for brakes from 34% to 84%. Fleets should measure accuracy on their own vehicles, in their own weather, in shadow mode before relying on it.
Pick one job, such as safety screening, damage liability or maintenance planning. Put the portal at one site where trucks already stop, like an exit gate. Agree a photographed defect list with maintenance, safety and claims before configuring anything. Run the system in shadow mode alongside existing inspections through bad weather, then connect flags to work orders and measure time from detection to repair.
Thinking about camera-based inspection for your fleet or yard? We can help you choose the first use case, write the defect list, assess vendors and plan the integration with your maintenance and transport systems.
What a native C trading-app study taught us about financial UX: tail latency, exact money, review-then-confirm flows, AI that proposes and code that decides.
Pairing a classifier with a frontier language model across 52 hotel email use cases cuts modelled AI API costs by 56% against an optimised Claude-only design.
Choosing where AI belongs is an engineering trade-off across accuracy, cost, complexity and risk. A CXO framework for humans, rules, machine learning and generative AI.