British Licence Plate Camera Fooled By Clothing

It’s a story that has caused consternation and mirth in equal measure amongst Brits, that the owners of a car in Surrey received a fine for driving in a bus lane miles away in Bath, when in fact the camera had been confused by the text on a sweater worn by a pedestrian. It seems the word “knitter” had been interpreted by the reader as “KN19 TER”, which as Brits will tell you follows the standard format for modern UK licence plate.

It gives us all a chance to have a good old laugh at the expense of the UK traffic authorities, but it raises some worthwhile points about the fallacy of relying on automatic cameras to dish out fines without human intervention. Except for the very oldest of cars, the British number plate follows an extremely distinctive high-contrast format of large black letters on a reflective white or yellow background, and since 2001 they have all had to use the same slightly authoritarian-named MANDATORY typeface. They are hardly the most challenging prospect for a number plate recognition system, but even when it makes mistakes the fact that ambiguous results aren’t subjected to a human checking stage before a fine is sent out seems rather chilling.

It also raise the prospect of yet more number-plate-related mischief, aside from SQL injection jokes and adversarial fashion, we can only imagine the havoc that could be caused were a protest group to launch a denial of service attack with activists sporting fake MANDATORY licence plates.

Header image, based on the work of ZElsb, CC BY-SA 4.0.

43 thoughts on “British Licence Plate Camera Fooled By Clothing

  1. In discussions about “AI” (haha, insane laughter from someone having used semi-automated self correcting statistical methods in software since the 1980s …) I always, almost every single time, have to suggest “humans checking results before blindly passing AI-output on for fully automated decision makings”.
    From “geeks” I get the eye. “Oh, that one again – AI is so good, it is soooo muuuuccchhhh beeedddaaa” (or are they saying “beta”?) ” than any human … we should eliminate the human error instead”.

    From normal people I get “wait, what, those results AREN’T CHECKED by humans?”

    It seems like there is a LOT of public education AND discussion needed. Still needed. Sigh. As if THAT would ever happen.

    1. Used correctly AI probably is better than a human, 100% of its maximum performance possibility every day all day. The downside is trying to use them on a less than constrained world – on your production line scenario the variation is so tightly controlled the AI is only presented with rather predictable inputs so won’t get it wrong much if at all while doing a job soo tedious a human would probably be loosing concentration…

      That is however the big problem AI – they are not as smart as even the stupidest human, so can’t deal with unexpected variables at all well – they won’t even realise its an unexpected event and thus have lower confidence in their result to bring an expert (human) into the loop – something this case really shows up, it looks like a number plate, its on the road, but a human would immediately know its not, its moving in the wrong way, its far too small and shaped incorrectly..

      This is something the AI can do – very good object detection with the right training set, but it needs the right training set – its got to be looking for more than just a number plate, so in this case probably wouldn’t trigger if it was also requiring the number it knows should be a Car/Motorbike/Van etc to look like one – could even flag up stolen plates potentially. But even then it will still see something similar as a high probability match, as its making decisions based on its training data, and the people who put that together will fail to account for something, because no human will think of everything that will trip the AI up if its not trained – so in the case of numberplate and car it might trigger on a toy car for instance – its got a number plate sticker, looks like a car the fact the scale is all wrong wasn’t something it considered – these odd edge cases are too obvious to us or to out of the ordinary to consider adding to the training data.

      Adversarial fashion is definitely going to be a headache for it as time goes on… Which personally I am fine with, if and when machines really become sentient they are probably more trustworthy with such jobs than a human, but as it stands now machines really should be kept as a tool for the human operators or in such controlled environments doing their very repetitive task – as that is what they are good for.

      1. OR…

        It was doing little more than OCR to read the plates, and correcting any issues with the system would require spending a few hundred grand the department that issues the fines just doesn’t have.
        In the UK, the department that fines people does not get to keep the money.

        Simple solution..
        Enclose a photo of the vehicle with the offending plate with every fine. Then when someone challenges the fine, as the vehicle in question appears to be a middle aged woman with text on her chest.. The fine is cancelled with a polite standard apology.
        Problem solved.

        1. Actually, it does take photos and those can be used to appeal against the fine. I was given a penalty notice once for driving in a bus only turn right lane in a town I had never visited (Reading, UK) but it had been snowing hard so there were no visible lane markings and the signs were unreadable from the dirt. I appealed using the photos provided by the local authority that showed a total white out, and the appeal was accepted.

          1. Reading is notorious for cameras catching you. Years ago they changed the direct of a one-way street one weekend, and fined 100s of people very fast – the police eventually intervened and reverted the change as it was getting dangerous.

        2. The wasted effort in man hours getting them cancelled and issuing apologies is going to add up to a pretty serious cost… Even more so if folks do start getting adversarial with the tech – which now its been so publicly borked is bound to happen…

          The sensible solution in that event, in that its probably cheaper and far more accurate would just be to ditch such cameras for the bus lanes and pay for a few more policefolk to stand/ride/drive around catching folks who abuse the system – most people are pretty law abiding even without big brother watching all the time, the point of the rules makes their lives safer/better as well… Up the penalty for getting caught as well perhaps, so the risk reward ratio tilts and it becomes not worth it for that reason too.

        3. That’s bullshit. Accusing someone of a *crime* based on laughably wrong evidence should require something more than just a “whoopsie.” Responding to specious criminal accusations is a cost on obviously innocent people, and an apology that doesn’t recompense them is unconscionable when the justification is so obviously laughable.

          1. Yes! Total BS, and incredibly harmful to society.

            “Problem solved.” But it was never solved, you just pushed the burden of your deficiencies on to innocent people who didn’t do anything wrong and should not be forced to shoulder the cost of your flawed system.

            While this is just license plates and tickets, other people are doing this with people’s faces and jail time. Guess what, these systems work fine for white folks, and has a terrible false positive rate for darker skin. What does the “Problem solved”, push the burden on to the wrongly accused victim look like then…

            Automated accusations of innocent people need remediation – something painful enough to the issuing group that they have an active interest in eliminating any false positives. Unfortunately we are so far from this…

    2. All UK traffic tickets have to be seen and approved by a human being by law. The cameras throw up false positives all the time, or there are mitigating circumstances behind the offence (Vehicle broken down in the other lane, road closure, etc) and many are discarded. What happened here is that the mandated human-in-the-loop wasn’t paying attention and authorised it to be passed on to the PCN system.

      In a previous life I coded the software that produces these images, as well as the systems that are used to process them. Trust me, staring at the output all day is mind-numbing at best and I personally cannot imagine a more boring job, except perhaps manually operating such a camera system in the first place. These things produce thousands of offences a day in a city like Bath, and someone (or really, many someones working in shifts) have to look amd approve or discard them.

  2. > but even when it makes mistakes the fact that ambiguous results aren’t subjected to a human checking stage before a fine is sent out seems rather chilling.

    Really ambiguous results are send for human processing on the system I worked on. But maybe this just wasn’t classified as ambiguous.

    Anyhow, it’s hardly the worst automated ticketing system error. And seeing the photo makes it obviously wrong and easy to correct.
    I do however, know of an incident with an average speed measurement system (two camera and measure the time it takes from cars to get from system 1 to system 2). Clocks drifted, 60.000 tickets where generated before the system was shutdown. And the company never sold another system like this again.

    Both systems where syncing by NTP, but when NTP went down on one system, nothing detected that and drift assured.

    1. Yeah, you could be that poor bastard in the US who chose “NULL” as a license plate, then got hit with every fine for every ticket that didn’t have a license plate attached, because of course the fine system just took the NULL from the database queries at face value.

      1. OT: I do like that someone put /dev/truck on rear of their vehicle…
        (I still might do it someday, but I want to have a Linux box connected to the Diagnostic Link Connector before I do that.)

  3. I would make a fake license plate that’s a copy of plate belonging to some official’s car. Like a mayor or representative of party I don’t like. I would put it on my car and do some reckless speeding and driving down bus lanes…

    In my country politicians routinely break the traffic regulations, like speed limits. A former prime minister crashed at least two state cars by her driver speeding way above the limit. So in my opinion it would be okay to use the copies politician’s plates. Unfortunately our plates are more complex than UK ones…

    1. Easier still, there are websites where you can order “show plates” that show a preview of how the resulting plate would look.

      Now that many UK cities are starting to charge for emissions with punitive daily charges just enter the details of your victim in one of the aforementioned sites and let the ANPR camera scan the image on a laptop. You could cause infinite disruption this way and I’m pretty sure there is no law against doing so.

    2. Which reminds me of the story of the man who gave a Mercedes to (the late) King Hussein of Jordan.
      In gratitude, the King gave his loyal subject a pair of Palace license plates for his own vehicle…i.e. no parking tickets ever!

  4. The systems used are actually quite good, it’s where the cost savings are made that gives terrible systems.
    Here – they could create a second level check, to filter out the oddities but, why bother, most just pay and don’t question…….
    Smart motorways, are a good idea. until you realise they’ve saved a small percentage by not bothering to install the amount of electronics needed to create the ‘smart’ bit.
    Yep, spend a hundred million quid removing the hard shoulder,
    save 5 million by not bothering to install enough cameras spaced close enough to detect vehicles,
    not having any redundancy in detecting vehicles.
    it’s idiotic short term cost savings over long term benefits

  5. AI is the very definition of the phrase “None of is as dumb as all of us”…
    AI can’t be any smarter than the dumbest person on the team that wrote it.
    Dunning-Kruger effect ensures that those making the decisions to adopt AI will always think it works great.

  6. The problem with the idea of activists wearing adversarial clothing en masse is that the UK parliament is fully sovereign, so all they have to do is pass a law banning it.

    By contrast, in the US, we have a constitution, and such activity would likely be protected by the first amendment.

  7. Here’s a funny license plate story.
    Here in Hollywood, many movie production companies utilize “Fake” license plates,
    made up to support the film script. Often, the prop master will use “Stick On” letters and numbers
    to create the “Fake”.
    The local newspaper, (L.A. Times) had a story about two LAPD officers arresting an “Actress”,
    due to her car having a plate with stickers on it. (It was used in the film)
    Locating her apartment, they broke in and hauled her outside without a chance to dress.
    Yes, she was starkers.
    They were convinced she was a hardened criminal preparing for some felony or other.
    The newspaper made quite a deal over the naked actress “Cuffed and Stuffed” into the back of a patrol car.

  8. It seems to me that it’d be an interesting idea to distribute low cost sweaters with various random license plate numbers in England. And then see if the system melts down…

  9. When I lived in Phoenix there were speeding cameras set up at different places on the freeways. They’d snap a picture of the driver and the car’s plate and then send a ticket in the mail. They had to match the driver’s face with the registered owner of the car from the driver’s license data base. Many got around it by having the wife register husband’s car and husband register the wife’s car. One guy got around it by wearing a gorilla mask when he drove. He generated a lot of press and the cops started following him around to try to catch him. IRIC, they were never able to prove it was him driving the car. The system was ultimately removed after court battles by people claiming they had a right to confront their accuser in court, and obviously that couldn’t happen. In the end it was determined that the contract for the system was issued after bribes were paid by the speeding camera company to some state politicians. AFAIK, there’s no longer any automated ticketing in Az now.

    1. Umm – the accuser is the state and the evidence is whatever was recorded by the camera. No different than if they find a fingerprint they say matches yours at the scene of a crime. You don’t get to throw it out because, you claim, you cannot confront the camera that took a picture of that fingerprint. As such, i doubt it was removed for that reason.

      It appears they are still doing automated ticketing in Arizona. and though Phoenix did decide to remove them

  10. If the camera well positioned and mounted to be very stable in extremely high winds, has a very high dynamic range (digital SLR cameras have about 8-11stops, the best cameras on the market have a dynamic range of about 15 stops, human eyes have about 24 stops) ideally augmented with a second near IR camera which is used with additional IR illumination from dusk to dawn and the camera(s) both would need to be very higher resolution (ANPR – Automatic Number Plate Recognition cameras are typically 4-12 megapixels, a single human eye is about ~576 megapixels) and there is no rain, sleet or snow then yes they might perform better than a human. But add some dust dirt and grimy build up in front of the lenses and suddenly the trained neural networks start to see things that are not there because they were trained with perfect data sets.

    But if you really wanted to mess with the system the easiest way would be a custom paint job where the entire car is just covered in numbers and letters in a similar font used for number plates. The thing about trained neural networks is that the edge cases where a human would laugh out loud if they saw the image that is where the system falls over because it is too far outside the standard training data. And it is impossible to train for the unexpected. It will probably get 99.99% right, but a human needs to be in the loop for the oddball ones.

    1. “It will probably get 99.99% right, but a human needs to be in the loop for the oddball ones.”

      Unfortunately, the government employee in the loop is often “the oddball one”.

    2. That resolution for the eye includes the pan-and=scan motion to apply central vision to the entire scene. For just central vision the effective resolution of the eye is far lower than that, and since reading the plate rather than detecting light is the goal, only the central vision matters. What is worse is the eye is subject to ghosting – as there is a rather large recover time from seeing the image to seeing a difference in the image, which allows movies and video to work with finite frame rates. A similar effect applies to the number of stops as the iris continuously modulates to adapt to the central vision illumination. Cameras can be used with the same adaptations but the comparison wouldn’t be as dramatic. More than that the brain really messes with what the eye reports – you don’t notice your own blind spots and not the time required to blink where you are entirely blind.

  11. I work with LPR cameras a lot and they can be spot on but they can still easily screw up and tag other things such as lettering from a sign, clothing and whatever it can see that has alphanumeric.

    Had one camera that when it saw a police car it would come up as PO11CE and uhaul trucks came up as U11AL. Keep in mind it would show the picture of what it saw.

    It’s easy to trick a camera during the day but at night it’s about 95% spot on since plates attend to be reflective.

    In Jacksonville Florida they had red light cameras and they were unmanned. When people found this out everyone started to challenge the tickets. After a year the city did away with them.

    1. >It’s easy to trick a camera during the day but at night it’s about 95% spot on since plates attend to be reflective.

      Back in the day I did a bit of work with some vendors who were developing LPR camera systems for several large contracts (customs between the US and Canada most specifically). In fact, they had the contracts for both sides.

      I asked if they exchanged data since it all seemed redundant, and was told that both data systems were so wildly different that it would be highly unlikely that they ever would, though the camera installations were similar. At that time (early 2000s) the cameras wouldn’t “see through” contaminants very well with the IR flashes they were using at the time (“A little smear of mud and you’re off the hook”), though this has likely been improved.

      Also, FWIW I drive a vehicle that I’ve slightly modified the plate with reflective tape to change a number to an alphanumeric character ( “7” to “?” if it matters). I’d love to see what the LPR systems make of that, but I’m not going to spend the money in fines to find out.

  12. The problem is whilst the fine might not be legal, neither is her number plate in the picture in the article.

    3d raised number plates are not legal in the UK.
    Whilst she should not have been fined for the bus lane, all the publicity she has courted should have resulted in her being fined for an illegal number plate.
    But as usual the police do nothing about it which is why we have a huge problem of dodgy plates in the UK.

    Two of the biggest scams going around the UK at the moment are raised 3d plate and smoked out plate – neither of which the police are really doing anything to prosecute for because the fine are too low.
    There is no good reason to mess with number plates.
    If the fine was £1000 AND surrendering your private plate then no one would be doing it.

    But the police, the very people responsible, are whining in the same press about ANPR not working correctly.

Leave a Reply

Please be kind and respectful to help make the comments section excellent. (Comment Policy)

This site uses Akismet to reduce spam. Learn how your comment data is processed.