Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Tesla Model 3 and speed limit signs

This site may earn commission on affiliate links.
I guess, like a lot of things Tesla are still developing, the speed sign recognition is basically just the first public version of that functionality. It's not reliable which is probably one reason why you have to tap the screen to set the AP limit rather than the car just doing it itself.

The part at 8:33 of this presentation explains some of the difficulties that they have with detecting "STOP" signs:

They can ask the cars to collect more examples to improve detection of things. So, presumably a similar situation exists for speed limit signs (e.g. need to better determine whether the sign applies to a side road, is the sign on the back of a truck, is it conditional based on the kind of vehicle or during school hours etc.)

It's a massive challenge, but it does look like they are aiming to be able to detect as many edge cases as possible (which they will absolutely need to do for FSD to actually drive itself.) And I agree it's quite insane to use AI to detect raindrops on the window, though I guess it means they don't have to interface with yet another kind of sensor? There's indications (from greentheonly on twitter) that they might even be developing their own parking sensors too so no doubt that will probably use some kind of AI as well!
 
Tesla software already interfaces with off-the-shelf parking sensors, so I don’t think it would be an issue for them to interface their software with off-the-shelf rain sensors and photoelectric ones for auto headlights.

In fact, since those technologies are so robust at this point I’d assume interfacing is child’s play in the grand scheme of things. They’re sensors so they just return a range of values that you respond to, which is considerably easier than performing some kind of raindrop detection function using one or more cameras. There is perhaps some logic in detecting far off rear lights, but again it’s surely something that is well and truly solved.

I’m not in the least bit surprised to hear that Tesla are looking into replacing another completely robust and proven system in ultrasonic sensors for parking, with some kind of camera based nonsense.

926DDC71-4227-469A-BD49-358817CC0D76.jpeg
 
I guess, like a lot of things Tesla are still developing, the speed sign recognition is basically just the first public version of that functionality. It's not reliable which is probably one reason why you have to tap the screen to set the AP limit rather than the car just doing it itself.

The part at 8:33 of this presentation explains some of the difficulties that they have with detecting "STOP" signs:

They can ask the cars to collect more examples to improve detection of things. So, presumably a similar situation exists for speed limit signs (e.g. need to better determine whether the sign applies to a side road, is the sign on the back of a truck, is it conditional based on the kind of vehicle or during school hours etc.)

It's a massive challenge, but it does look like they are aiming to be able to detect as many edge cases as possible (which they will absolutely need to do for FSD to actually drive itself.) And I agree it's quite insane to use AI to detect raindrops on the window, though I guess it means they don't have to interface with yet another kind of sensor? There's indications (from greentheonly on twitter) that they might even be developing their own parking sensors too so no doubt that will probably use some kind of AI as well!

call me cynical but when you repeatedly miss your goals you either confess you’ve leader has been overly optimistic or you make out the problem is much harder than anyone could imagine. Speed limit recognition was pretty good on AP1 cars using tech developed nearly 10 years ago which Tesla haven’t yet been able to match. Stop signs the same really.
 
call me cynical but when you repeatedly miss your goals you either confess you’ve leader has been overly optimistic or you make out the problem is much harder than anyone could imagine. Speed limit recognition was pretty good on AP1 cars using tech developed nearly 10 years ago which Tesla haven’t yet been able to match. Stop signs the same really.

I can answer that one maybe. The issue is Mobileye own the patents on the efficient way of reading the signs. I haven't read the patents, but I suspect they're looking for a specific shape, and then literally reading the numbers inside. Somewhat like traditional OCR. Tesla don't like licensing, and they especially don't want to do any licensing with Mobileye because they fell out.

For Tesla, they're using AI/ML instead of optimised custom code. It's possible to do it this way (obviously!) but it's not the most reliabile or efficient method. I suspect they're going for a quick bit of pattern recognition. So for example, while Mobileye will work more like you do - recognise a sign, read in the number. Tesla will be looking at the sign as a whole. Imagine you came from another planet and don't know what 0-9 are. But what you do know is if there's a round sign with red borders and a black '30' in it, that means 30mph limit in your own units / characters. There's only a finite number of discrete numbers so you just have to memorise the common speed limit signs and you're done - except for the one-off signs but they're rare. Tesla could of course be doing much more, they could have pattern recognition on the 0-9 symbols too and re-create the sign internally. If Tesla ever recognises a sign with an odd speed like '23' then we know they're matching down to the characters. If they only ever recognise the 'standard' speed limits, then it's reading in the sign as a whole.

Basically it's like the windscreen wipers. They're using a really difficult and complicated method because they don't want to use 3rd party licenses and tech wherever possible. They can't re-create the 3rd party tech because of said licenses / patents.
 
I can answer that one maybe. The issue is Mobileye own the patents on the efficient way of reading the signs. I haven't read the patents, but I suspect they're looking for a specific shape, and then literally reading the numbers inside. Somewhat like traditional OCR. Tesla don't like licensing, and they especially don't want to do any licensing with Mobileye because they fell out....
I guess this approach may be a good way to recognise non numeric traffic signs of which there are many, in fact most signs that are not speed limit signs.
 
I'm aware Mobileye hold the patent, but you don't half make life difficult for yourself if you refuse to accept these things, and end up making a pigs ear of it trying to make a square wheel. Same goes with the map data, they're running 2019 maps because the 2020 version was rubbish and they won't use a decent map vendor. The live google map in the car is also a facade, it doesn't do anything in the car other than for a backcloth for the car position, its certainly not involved in the Nav routing. As the world changes the data becomes more out of date, new speed limits and junctions aren't picked up and reflected in the car, and mire mistakes get made.
 
....Same goes with the map data, they're running 2019 maps because the 2020 version was rubbish and they won't use a decent map vendor. The live google map in the car is also a facade, it doesn't do anything in the car other than for a backcloth for the car position, its certainly not involved in the Nav routing....
Thanks - I didn't know that it does not really "use" google maps. Shame.
 
There really isn't anything wrong with how Tesla detect speed signs, it's just as accurate as the Mobileye solution in my other car. Neither are totally unfoolable, but then I am not either with some strange layout of signs. I'm sure they will add motorway speed detection at some point, but it's hardly the biggest issue as drivers have got plenty of time to be aware given Autopilot is generally doing the driving on motorways.

Thanks - I didn't know that it does not really "use" google maps. Shame.
Using Google maps to route requires an internet connection, so wouldn't really be a very reliable solution for a car. You can see this in Polestar where it can't route if there is no internet.

Tesla uses offline maps from OpenStreetMap to route, just overlays Google maps or satellite if data is available. It's updated automatically but isn't always totally up to date. Yes they could buy more reliably updated map data from Here and the like, but you only see an issue with really new roads.
 
  • Helpful
Reactions: justinhow
If Tesla manage to do everything themselves in the weird way that they are, aside from not needing to be concerned with licensing or patents they have full control of their implementation so can do whatever they like - be that to develop it further or to license it along with the rest of their AI stuff to other companies. I doubt they'd be able to enhance (or break...) things more if they were just using someone else's tech.

The current situation is admittedly far from ideal and when you pick any individual piece of functionality there are better, working solutions out there already for most of those things (auto wipers, blind spot detection, reversing cameras, lane departure warning) and things we don't even have (speed camera warnings.) Those all probably work perfectly fine, albeit probably a bit more tricky to update (if you even need to.)

I'd say it's probably that they want to maximise income by not having to pay other people for the tech, but then again I'd imagine they're spending a fortune on developing all of this stuff themselves (from the offline AI training systems to designing and manufacturing the FSD computer etc.)

The main advantage (for users?) of Tesla doing it all themselves is being able to push out updates and collect data to improve the AI. The onboard computer is essentially one "brain" (well, actually two) dealing with the various cameras and sensors the car has. As this becomes more capable, the car should get better at all of the things it does, with minimal or no hardware changes. Free software updates bring new stuff - though it would be really nice if they didn't come with bugs too.

I remember the weather was peeing it down for seemingly the entire month after I collected my car back in November 2019. The auto wipers were utterly useless. They have improved significantly since then, though there are occasions when I need to give it a nudge still.

FSD beta in the US has its quirks still but it's pretty impressive how far things have come.

On the subject of speed signs though: I wonder which system will best deal with this?
Mystery of Ainsley Harriott faces on Derbyshire road signs
 
If Tesla manage to do everything themselves in the weird way that they are, aside from not needing to be concerned with licensing or patents they have full control of their implementation so can do whatever they like - be that to develop it further or to license it along with the rest of their AI stuff to other companies. I doubt they'd be able to enhance (or break...) things more if they were just using someone else's tech.

The current situation is admittedly far from ideal and when you pick any individual piece of functionality there are better, working solutions out there already for most of those things (auto wipers, blind spot detection, reversing cameras, lane departure warning) and things we don't even have (speed camera warnings.) Those all probably work perfectly fine, albeit probably a bit more tricky to update (if you even need to.)

I'd say it's probably that they want to maximise income by not having to pay other people for the tech, but then again I'd imagine they're spending a fortune on developing all of this stuff themselves (from the offline AI training systems to designing and manufacturing the FSD computer etc.)

The main advantage (for users?) of Tesla doing it all themselves is being able to push out updates and collect data to improve the AI. The onboard computer is essentially one "brain" (well, actually two) dealing with the various cameras and sensors the car has. As this becomes more capable, the car should get better at all of the things it does, with minimal or no hardware changes. Free software updates bring new stuff - though it would be really nice if they didn't come with bugs too.

I remember the weather was peeing it down for seemingly the entire month after I collected my car back in November 2019. The auto wipers were utterly useless. They have improved significantly since then, though there are occasions when I need to give it a nudge still.

FSD beta in the US has its quirks still but it's pretty impressive how far things have come.

On the subject of speed signs though: I wonder which system will best deal with this?
Mystery of Ainsley Harriott faces on Derbyshire road signs
The problem with this is they AREN'T managing to do everything for themselves. They are ignoring something that works in order to try and develop another solution. They are failing. I wouldn't mind if these things worked but they just don't. If it's a new problem then fine but to be expending all this energy on something that is already fixed is crazy.

It'll turn people off. Me, for example, there is no way my next car will be a Tesla until functions that are standard on other cars also work on a Tesla.
 
Using Google maps to route requires an internet connection, so wouldn't really be a very reliable solution for a car. You can see this in Polestar where it can't route if there is no internet.

Tesla uses offline maps from OpenStreetMap to route, just overlays Google maps or satellite if data is available. It's updated automatically but isn't always totally up to date. Yes they could buy more reliably updated map data from Here and the like, but you only see an issue with really new roads.

You still requires an internet connection to setup navigation. You cannot even set a destination without the car being online. But once navigation is setup, I've never noticed any issues with it not being able to navigate.

Worth pointing out that depending on which model car you get, and a lesser extent, how you pay for it, you have to pay ~£10/month to get live traffic visualisation (includes other features too) after the 1 or 12 month 'premium connectivity' trial. However navigation uses a different behind the scenes method for traffic rerouting so is always available even if you cannot see why.
 
I guess this approach may be a good way to recognise non numeric traffic signs of which there are many, in fact most signs that are not speed limit signs.

That's a very good point that I didn't think about. For FSD they need to know all signs. That makes using an ML solution much more useful since the speed limit signs are just 'another sign' they add to the process and not a special case. Not having to get developers to write custom code for every sign in the world is a big deal.
 
Mobileye etc presumably have to solve the same problem, though.

That being said I can understand the logic of Tesla going their own way on that side of things, simply because it can be argued that it’s a use case for FSD.

What I don’t agree with is Tesla reinventing the wheel and coming up with a square with stuff like rain sensors, photoelectric sensors (auto headlights), etc. No customer cares that Tesla make the hardware for those things, they just care that it works reliably, which it clearly doesn’t.

Every other manufacturer just uses the off the shelf components and interfaces their systems with them. They are not “in development” or subject to future updates, they are well refined, proven, and complete.
 
Google maps has a perfectly servicable offline mode.
Try it on your phone, it only support routing for areas you have chosen to take offline. Yes there is a large scale view of roads in other locations, but no routing.

Also check out some videos of Polestar2, same experience.

Kind of immaterial anyway, these are features of the Google Maps Android application, Tesla is not running Android, can't use the apps. They are using the public APIs to retrieve map tiles.
 
I navigated from Calais all the way to Luxembourg offline on google maps without having to choose anything.

Once you get out of the cached area it has a bunch of low resolution maps - pretty much main roads only - that you can navigate around. I was surprised how well it worked.. it got me to my destination in a country I'd never visited before with zero access to contact anyone to get directions..

(Three's fault - it turns out when you buy the roaming package it doesn't automatically enable roaming - you have to call them - which isn't explained and I didn't find out until I was out of the UK. Didn't even get my money for the package back...)
 
I navigated from Calais all the way to Luxembourg offline on google maps without having to choose anything.

Once you get out of the cached area it has a bunch of low resolution maps - pretty much main roads only - that you can navigate around. I was surprised how well it worked.. it got me to my destination in a country I'd never visited before with zero access to contact anyone to get directions..

(Three's fault - it turns out when you buy the roaming package it doesn't automatically enable roaming - you have to call them - which isn't explained and I didn't find out until I was out of the UK. Didn't even get my money for the package back...)
I guess that's because it had defined the route when you were online.
 
Three's fault - it turns out when you buy the roaming package it doesn't automatically enable roaming - you have to call them - which isn't explained and I didn't find out until I was out of the UK. Didn't even get my money for the package back...)

Are you saying that if you take an M3 abroad the connectivity package doesn’t work without contacting Tesla to get it turned on in roaming mode first? I thought I had understood it was already in roaming mode in the U.K. with a SIM registered in Netherlands or Spain or somewhere.