More than two million motorists could be wrongly identified for fines by Ulez or speed cameras every day, a Government commissioner has warned ministers.
Professor Fraser Sampson, the surveillance camera commissioner, said the ANPR (Automatic Number Plate Recognition) camera network’s three per cent error rate for reading car number plates meant there were “significant risks” of penalty notices being wrongly issued to innocent motorists.
3% error rate? That’s appalling. How did anyone roll out a tech that shite?
Because these are the same numbskulls who think you can power a modern economy with unicorn farts and windmills.
They are politicians: therefore a thick as concrete and proud of their innumeracy.
I look forward to millions of challenged ‘offences’ being raised every day. Let’s see the politicians weasel out of this. Did anyone say ‘summary justice’?
Steve: I hope your lions are hungry. They’re needed.
Steve’s lions are always hungry because he feeds them on tofu.
Hang on, this isn’t new to ULEZ. The existing congestion charging zone, est. 2003, would have had the same problem. The smaller ULEZ, est. 2019, would have had the same problem. The police also use ANPR for checking that cars are taxed & insured. Did nobody notice the 3% error rate before, or is it just made-up?
Re-train the AI model? More HF in the RL?
Maybe it believes BMW drivers are gorillas.
3% error rate? That’s appalling. How did anyone roll out a tech that shite?
You reckon NHS (Envy of the World) diagnosis error rates are better?
“How did anyone roll out a tech that shite?”
A 3% error rate is a dead cert compared to facial recognition, which is apparently used by all police forces now, even if they don’t admit it. Quite literally, showing your face in public puts you at risk of a criminal charge.
I’m not sure where he’s getting 3% from, because OpenALPR is over 99% accurate, and OpenALPR is more of an off-the-shelf thing with a single algorithm trained with license plates for each country. I’m sure there are better, more expensive ones than that. Perhaps he’s got confused between detection and accuracy: can it see a plate and see digits, or is it getting the digits right?
I mean, he’s self-evidently wrong because how often do you hear anyone say they got a ULEZ letter when they weren’t in London?
“the surveillance camera commissioner” – how the hell did we get to here?
Ducky McDuckface: “Maybe it believes BMW drivers are gorillas”
Then it is indeed wrong. BMW drivers are bonobos. The gorillas are in Audis.
thats ok because 10% of the numberplates are fakes, clones or deliberately obscured anyway. Those would mostly be on the few vehicles that actually attract the tax.
Stupid tax, insanely expensive to collect, taxes the poor, encourages avoidance, doesnt tax tyre pollution which is worse from modern heavy cars, especially EV’s.
So it’s not really about the tax is it, big brother is watching you!
The 3% error rate is most probably the number of plates it fails to read due to them being dirty or the lighting conditions making it tricky. If it was 3% being identified as the wrong number it would still not lead to millions of people being fined as the car number space is sparce because there are lots and lots of possible numbers not assigned to any car.
If it were a major problem we would have heard the screams of indignation by now. As always real world evidence beats the theory of some professor as interpreted by the Telegraph.
What about alignment/field for cameras, you drive close to but not into the zone and it charges you? How clever is it if there are reflective surfaces in its field of view?
@WB
One always has to look at the application not the name-plate (As per windfarms)
I’ve been using OCR for years. Given ideal circumstances, it’s 100% accurate. But in practise there’s a varying error rate according to the quality of the input.
I’d go along with Andy F. There’s something like 80 million letter combinations for each 6 month numeral* issue. So the chances of the ULEZ system hitting an issued registration sequence in error are fairly low.
*Curious how motor industry lobbying kept that after you moved on from year letters. There’s not the slightest reason for it & hasn’t been for years. The police know when your car was first registered as they know whether it’s taxed of insured. And as I found out to my cost & inconvenience, even the Spanish police know it. (DVLC failed to do the direct debit). It’s solely so you can inform the neighbours you’ve again bought a new car. Is the UK the only country in Europe does this?
@BiS
It’s actually quite handy when buying a second-hand car: you can trivially tell how old it is, so the unscrupulous can’t lie about the car’s age in the advert.
BIS,
“I’ve been using OCR for years. Given ideal circumstances, it’s 100% accurate. But in practise there’s a varying error rate according to the quality of the input.”
It’s similar but different. On the one hand, you have weather problems, but on the other hand you have a moving vehicle, so the system gets many shots to detect it at slightly different positions. And a license plate has a certain structure to it, fixed font size and font type. Like if you only get a certain number of characters, you know it’s wrong. It’s closer to the old thing of banks scanning the special characters on the bottom of cheques in that regard.
There’s something like 80 million letter combinations for each 6 month numeral* issue. So the chances of the ULEZ system hitting an issued registration sequence in error are fairly low.
If ANPR was generating completely random 7-figure alphanumeric strings (in a valid format for a UK plate), then you’d be correct. But what happens in reality is that an M on a valid plate gets misidentified as a N, so AB23 KLM becomes AB23 KLN. The odds of the incorrect value having been assigned to another vehicle are then pretty good.
How do you get to 80 million? Even with the full 26 letters, I make that ~12 million, and then of course some letters aren’t used.
You can tell when the error rate was considered low enough – it was when people were allowed to register for automatic charging based on ANPR rather than having to volunteer payment with ANPR used as a threat.
ANPR cameras use infra-red, usually with their own illuminators, so they get a far clearer picture of the text than you might expect. See the picture here: https://www.ea-group.co.uk/security-systems/anpr-automatic-number-plate-recognition/ though that’s actually at the low end of quality. And the camera usually also takes a visible light picture which can be manually checked – which can catch people whose number plates are deliberately unreadable in IR light as a human operator can see the plate as well as any other features of the vehicle that might identify it.
But note that the story is hopelessly garbled. The two million motorists is derived from 80 million plates read per day. Even with a 3% error rate, that can only mean 2.4 million motorists fined if that many fines are actually issued. Many of the errors will be obviously wrong (no such plate exists), some will falsely detect a plate entitled to be there etc. And it is clearly impossible to keep fining 2.4 million motorists a day as the first year would see 876 million motorists fined, which is vastly exceeding the number of motorists in the country. There is a difference between a fine and a driver. If there were 2.4 million fines per day – whether valid or not – that’s 17.5 per licensed driver. Or one every three weeks per driver. That is hugely more than anything plausible.