Electric car for £9500?

[coppice]

Just on the way back from London my ID.3 perceived "something" (I've no idea what) on the M25 as a 100 mph speed limit sign and set the ACC to 100 mph.  I would have thought the software would be smart enough to know the maximum legal limits for certain countries...

I am tempted to attach an OBD11 and code out a lot of these "safety" functions, they're not ready for prime time. The worst one is lane centering on badly marked rural roads, it likes to ditchfind, so you have to be quite firm with the steering wheel to override it.

A few months ago driving north on the M! there was a long section of roadworks limited to 50MPH, with many 50 limit signs. Every one my car passed read as 60. They were all clean and clearly printed. The car is normally pretty reliable reading those numbers. The weather was good, and there were no strong reflections which might have affected it. Heaven knows what gets into these systems. Most cars use a mix of limit signs and mapping for their dashboard display of the speed limit. My car shows a faint queued up next limit as I approach a new section of road, based only on mapping (it often comes up before the sign is visible). The actual sign reading seems to be mostly for extra limiting in bad weather or due to road works. Its strange that your car would go above the normal limit, as I assume it will also have that overall mapping.

[nctnico]

Just on the way back from London my ID.3 perceived "something" (I've no idea what) on the M25 as a 100 mph speed limit sign and set the ACC to 100 mph.  I would have thought the software would be smart enough to know the maximum legal limits for certain countries...

I am tempted to attach an OBD11 and code out a lot of these "safety" functions, they're not ready for prime time. The worst one is lane centering on badly marked rural roads, it likes to ditchfind, so you have to be quite firm with the steering wheel to override it.

Volkswagen... what do you expect? A couple of years ago I rented a Peugeot which also had automatic speed limit reading. It failed pretty bad while driving in France. Go figure.

[tom66]

Volkswagen... what do you expect? A couple of years ago I rented a Peugeot which also had automatic speed limit reading. It failed pretty bad while driving in France. Go figure.

Well, I'm sure the camera system is made by Mobileye or something, VW just integrate it.

I've got a Vauxhall Mokka loaner from Enterprise (no, it wasn't my choice) whilst they repair the bumper on my car - it seems to misread most of the variable speed limit signs around here too.  It also assumes that after 1-2 miles of not seeing any other road sign that it should hide the speed limit in case it has changed, but this isn't very useful as many roads have only occasional repeaters.

It's one of those features that sounds good on paper but the reality is it kinda sucks and it needs more work.

[Veteran68]

Well interestingly enough, the speed limit reader on my Audi works very reliably. Surprisingly so. Sometimes it can even determine dynamic zones with different speed limits based on time of day and flashing or not flashing lights (like schools zones). That part is really hit or miss, but for static signs it rarely if ever gets it wrong.

[Siwastaja]

What's the peak and mean LIDAR output power? How does that compare to eye-safe limits/

I don't know exactly in numbers, but what I do know is that industrial LIDARs that have been approved to use in the middle of people (e.g. workers in factory) and where direct eye contact just centimeters apart does no harm, have been demonstrated in use in automotive conditions (e.g., rain) and work fine, too, to distances of at least 100 meters. The problem is mostly in price (and all those nasty engineering details like keeping the sensor glass clean, but apparently they somehow manage to do that with cameras, too.

[tggzzz]

What's the peak and mean LIDAR output power? How does that compare to eye-safe limits/

I don't know exactly in numbers, but what I do know is that industrial LIDARs that have been approved to use in the middle of people (e.g. workers in factory) and where direct eye contact just centimeters apart does no harm, have been demonstrated in use in automotive conditions (e.g., rain) and work fine, too, to distances of at least 100 meters. The problem is mostly in price (and all those nasty engineering details like keeping the sensor glass clean, but apparently they somehow manage to do that with cameras, too.

I presume the required power is the same as for RADAR, i.e  P=range⁴ See p16 of https://www.ll.mit.edu/sites/default/files/outreach/doc/2018-07/lecture%202.pdf Thus if, say, 1mW is sufficient for 1m then 10W is required for 10m and 100kW for 100m. Ouch.

But that's peak power. In slowly changing environments the environment won't need to be "sampled" often, so the mean power could be much lower. But streets are fast-changing, so the mean power would be higher.

Overall, "numbers, not adjectives" are necessary

[Siwastaja]

Overall, "numbers, not adjectives" are necessary

On the other hand, you can observe the fact that an apple falls from a tree towards the planet, and an airplane takes off and flies, without knowing the value for g, or how the wing aerodynamics are calculated. I have made a similar observation about LIDARs. They exist in full-autonomous car prototypes I have personally seen demonstrated, using the exact same LIDARs sold for industrial environments approved working close to humans, and I have personally seen these LIDARs map obstacles in automotive environment. Thus, I don't need to know what the exact power is, and how eye safety is exactly proven; this would require deep dive into the relevant legislation and look at the LIDAR manufacturer's (e.g. SICK).

In the end, devil is in details and I don't exactly know how well modern state-of-the-art LIDARs perform in presence of thick rain/snow, but my educated guess is, they do work, unlike cameras. One distinctive feature of a modern LIDAR is, for given direction, it can represent multiple echoes (distances).

[tggzzz]

Overall, "numbers, not adjectives" are necessary

On the other hand, you can observe the fact that an apple falls from a tree towards the planet, and an airplane takes off and flies, without knowing the value for g, or how the wing aerodynamics are calculated. I have made a similar observation about LIDARs. They exist in full-autonomous car prototypes I have personally seen demonstrated, using the exact same LIDARs sold for industrial environments approved working close to humans, and I have personally seen these LIDARs map obstacles in automotive environment. Thus, I don't need to know what the exact power is, and how eye safety is exactly proven; this would require deep dive into the relevant legislation and look at the LIDAR manufacturer's (e.g. SICK).

In the end, devil is in details and I don't exactly know how well modern state-of-the-art LIDARs perform in presence of thick rain/snow, but my educated guess is, they do work, unlike cameras. One distinctive feature of a modern LIDAR is, for given direction, it can represent multiple echoes (distances).

Points taken, but how wings work is an exceptionally bad example: most explanations - especially in school textbooks - are simply wrong!

I think you'll find there are cases where LIDAR is better than cameras, and vice versa. "Sensor fusion" has long been an interesting and important topic.

W.r.t. multiple echoes/distances, that introduces ambiguities as well as resolves (ho ho) them.

[tom66]

What I don't fully appreciate with LIDAR is, when a substantial number of cars are running LIDAR, will there not be an issue with interference between similar LIDARs using similar wavelengths?  I presume there is some kind of CDMA or similar on the light pattern, but still could imagine eventually the additional light will reduce the sensitivity too much as it becomes difficult to separate 'your' code from 'their' code.  But maybe this limit isn't reached in practice.

[coppice]

What I don't fully appreciate with LIDAR is, when a substantial number of cars are running LIDAR, will there not be an issue with interference between similar LIDARs using similar wavelengths?  I presume there is some kind of CDMA or similar on the light pattern, but still could imagine eventually the additional light will reduce the sensitivity too much as it becomes difficult to separate 'your' code from 'their' code.  But maybe this limit isn't reached in practice.

LIDAR is highly directional, so the conflicts are probably not that big an issue. What happens as more and more cars have ultrasonic and radar sensors, with quite broad beams? The ultrasonic stuff is only going to mess things up at low speed for other cars close in, but radar cruise control illuminates quite a lot of cars on a busy highway.

[Siwastaja]

I think you'll find there are cases where LIDAR is better than cameras, and vice versa. "Sensor fusion" has long been an interesting and important topic.

Spot on. I don't believe in the ability of doing reliable obstacle detection and emergency braking without combining at least two very different types of data source, vision and some sort of distance measurement (stereo vision does not count); human-like vision plus a "superhuman" one. It's sad to see how tech companies see all the buzz in trying to replicate very complex human systems like gigapixel vision and gazillion-neuron NNs, when they could assist these worse-than-human replicas with some superhuman skills like ultrasonic, radar, LIDAR, etc., pretty easily!

[nctnico]

I think you'll find there are cases where LIDAR is better than cameras, and vice versa. "Sensor fusion" has long been an interesting and important topic.

Spot on. I don't believe in the ability of doing reliable obstacle detection and emergency braking without combining at least two very different types of data source, vision and some sort of distance measurement (stereo vision does not count); human-like vision plus a "superhuman" one. It's sad to see how tech companies see all the buzz in trying to replicate very complex human systems like gigapixel vision and gazillion-neuron NNs, when they could assist these worse-than-human replicas with some superhuman skills like ultrasonic, radar, LIDAR, etc., pretty easily!

Throwing more sensors at a problem is not a solution in itself. The key problem to solve is to determine when the sensor data is correct and when not. That will take most of the computational power.

For example: In the early 2000's I worked at a research institute where one of my ongoing projects involved an instrumented car which was used for research into how people drive a car. This had a lidar for measuring distance between the car in front and a scanning sensor to measure the distance to the line on the side of the road. When I first started people had to sort through the data manually to get rid of pieces of data that where invalid. That was a lot of work so I automated it through software. One of the problems with the Lidar was that it would pick up trees or signs when driving through corners so I set a limit on the valid distance based on speed / steer rotation. The line detector was also good at detecting puddles with water. So I created an algorithm that looked for a line that was a) wide enough and b) followed the general distance between the car and the line. A car won't jump 25cm left/right in less than 0.1 seconds.

[coppice]

Throwing more sensors at a problem is not a solution in itself. The key problem to solve is to determine when the sensor data is correct and when not. That will take most of the computational power.

Why do you think that? Do you have an reference models to base that on? Everything I've ever seen where results of sensing are solid and reliable use some form of sensor fusion specifically to resolve which data is correct. Biological systems do this to an extreme. Our own senses are very easy to fool in isolation, while the fused set works quite well, Most illusions are based on the senses not being able to fuse as they would otherwise do. There is hardly anything we get consistently right with just one sense. Good solid science and engineering work almost always relies on looking at things in multiple ways to resolve ambiguities.

[nctnico]

Throwing more sensors at a problem is not a solution in itself. The key problem to solve is to determine when the sensor data is correct and when not. That will take most of the computational power.

Why do you think that? Do you have an reference models to base that on? Everything I've ever seen where results of sensing are solid and reliable use some form of sensor fusion specifically to resolve which data is correct. Biological systems do this to an extreme. Our own senses are very easy to fool in isolation, while the fused set works quite well, Most illusions are based on the senses not being able to fuse as they would otherwise do. There is hardly anything we get consistently right with just one sense. Good solid science and engineering work almost always relies on looking at things in multiple ways to resolve ambiguities.

That is exactly what I wrote...

[tszaboo]

Just on the way back from London my ID.3 perceived "something" (I've no idea what) on the M25 as a 100 mph speed limit sign and set the ACC to 100 mph.  I would have thought the software would be smart enough to know the maximum legal limits for certain countries...

I am tempted to attach an OBD11 and code out a lot of these "safety" functions, they're not ready for prime time. The worst one is lane centering on badly marked rural roads, it likes to ditchfind, so you have to be quite firm with the steering wheel to override it.

Volkswagen... what do you expect? A couple of years ago I rented a Peugeot which also had automatic speed limit reading. It failed pretty bad while driving in France. Go figure.

People reported this driving a Lexus.

[coppice]

Throwing more sensors at a problem is not a solution in itself. The key problem to solve is to determine when the sensor data is correct and when not. That will take most of the computational power.

Why do you think that? Do you have an reference models to base that on? Everything I've ever seen where results of sensing are solid and reliable use some form of sensor fusion specifically to resolve which data is correct. Biological systems do this to an extreme. Our own senses are very easy to fool in isolation, while the fused set works quite well, Most illusions are based on the senses not being able to fuse as they would otherwise do. There is hardly anything we get consistently right with just one sense. Good solid science and engineering work almost always relies on looking at things in multiple ways to resolve ambiguities.

That is exactly what I wrote...

It seems to be the opposite of what you wrote. You seemed to be emphasising that well thought out algorithms can do wonders without adding more sensors. The problem is this only goes so far. Our brain is using lots of the kind of heuristic algorithms you described, and it fails a lot. The flexibility needed to deal with a complex and unpredictable world needs multiple sources of information to resolve things.

[mikeselectricstuff]

What I don't fully appreciate with LIDAR is, when a substantial number of cars are running LIDAR, will there not be an issue with interference between similar LIDARs using similar wavelengths?  I presume there is some kind of CDMA or similar on the light pattern, but still could imagine eventually the additional light will reduce the sensitivity too much as it becomes difficult to separate 'your' code from 'their' code.  But maybe this limit isn't reached in practice.

AIUI the duty cycle (in a particular direction) is very low - you want high peak power for optimum detectability, so this probably gives a reasonable head start in interference rejection. On top of that you can apply a lot of "reality checks" based on frame-to-frame changes based on likely velocities of real-world objects

[Siwastaja]

Throwing more sensors at a problem is not a solution in itself. The key problem to solve is to determine when the sensor data is correct and when not. That will take most of the computational power.

Exactly. And different types of sensors, with different strengths and weaknesses, enables reliable sensor fusion. Algorithms are still not simple, but at least it's possible. I have zero trust on camera-only neural network stuff for anything serious like safety-critical obstacle avoidance. Adding a LIDAR does not magically make it happen, but it makes it possible. Computation is still a big risk item though.