odysseus2000 wrote:One of the long running discussions surrounding self driving car sales is whether Lidar, as used by folk like Google, is better than a visual system as use does by Tesla, but what is Lidar, how does it work & what are its limitations? Thus is an interesting video that covers a simple scanning Lidar system now available for around $100 & shows its strengths & weakness & makes coments re full 3d systems etc, well worth a watch imho:
https://youtu.be/AERfie9wlWMRegards,
Just watched it.
What a load of twaddle.
All it is, is a video examining one, cheap lidar sensor, and considering whether it would be suitable as the only single device by which you could drive a toy car.
It has nothing to do with "as used by folk like Google".
It has nothing to do with whether it's 'better than a visual' - that's a gross misrepresentation of the reality when it comes to self driving cars.
And the points the narrator makes do not stand up to scrutiny either.
If he kept the discussion to hobbyist toying around - which is in essence what he's doing - then it would be fine. A little project with your kids to get them interested in the technology, fine. But to suggest - as even the description on the video itself does - that this is somehow transferrable to real self driving cars, nah ... that's let's say 'naïve'.
To address these, and other problems with the video...
Point 1
No real-world-for-humans-and-amongst-humans proper autonomous car is even being considered to be operated by lidar alone. To claim (or at least imply) Lidar is not suitable or appropriate at all for self driving cars simply because it cannot do everything in one single device, is completely preposterous. No (serious) manufacturer is trying to make an autonomous vehicle using lidar alone.
All technologies have limitions, that's why most companies use a combination of technologies in their self driving vehicles.
Much as I'm an admirer of the state of the art in visual recognition (see earlier posts I've made on this board) - it has certainly undergone a revolution recently and self driving cars certainly are not viable without it - but to rely solely on vision would be as mad as relying solely on lidar (and yes I am aware that one company is trying to rely solely on vision - I hope that doesn't succeed in becoming the standard - computer vision is now immensly powerful using deep learning convolutional neural nets, but it isn't completely foolproof - any more than the lidar sensor is completely foolproof.).
But to imply that you might consider using lidar alone for a real world self driving car is ridiculous. How is lidar going to see what state traffic lights are? How is lidar going to recognise a policeman directing traffic? How is lidar going to see speed signs, road work signs, diversion signs, road closed signs?
How is lidar going to identify an object so that it can predict how it might move? How would it recognise a car, bike, lamppost, garbage, leaves, cats, dogs, children, prams, etc.
It is quite simply preposterous to assume that someone might consider lidar without an associated vision system.
What lidar does give you is a more concrete feedback on distances - which can then be combined into the visual data to double check the spatial model being created. We all know how easy it is to camouflage things visually. Lidar gives a second source of information about objects that isn't prone to issues of visual camouflage.
Lidar compliments visual recognition. It cannot replace it.
Point 2
He claims you don't need to see behing when driving forwards. Utter twaddle. Any serious self driving system will need to pay attention to the traffic behind. How do you know when it's safe to make a lane change? How do you know there isn't a cyclist overtaking / undertaking?
I've seen seen already on a number of real world cars when I'm cycling, that they now have warning sensors that flash a light in the left wing mirror to warn the driver when a cyclist is approaching from behind / passing on the left. Watching what's coming up behind is critical to being able to drive safely.
Point 3
He claimed - even using stealth bombers to show how impressive his argument is - that it can only see walls it's looking straight at, rather than at an angle. But if you look at his demo you could see that it was actually perfectly able to see the wall into the distance.
If you look closely at the video at 5:51
https://youtu.be/AERfie9wlWM?t=351 where he claims to be showing it not seeing the wall, the picture on screen tells a different message. If you pause it, and look closely, you can see dots practically all the way to the top of the screen.
What you are seeing is not that it isn't seeing the wall, what you are seeing is that as the wall goes further away each angle the lidar rotates through is covering a much greater length of wall, so the dots stretch out when drawn on the display.
That is what you are seeing - less dots per distance of wall, due to more distance of wall per angle of lidar - you are not seeing it fail to see the wall.
Point 4
And he does admit this near the end - there are lidars that aren't so limited to the 2d plane. There are lidars that do cover multiple heights. Which kind of undermines much of his criticism of that lidar when looking at lidar in general for self driving cars.
So in essence, the whole video is pointless - twaddle.
He's taken one of the most budget lidars around, shown that it isn't great (wow, the cheapest lidar doesn't do everything ... and in other news)...
And then further tries to show "the ugly truth" about lidar on the (false) basis that someone might actually not only consider using only lidar for a self driving car, but even more foolishly would even consider only using a single one of these cheapest lidar as the sole sensor in a self driving car.
As one of the comments illustrates...
"Very interesting. It covers perfectly the problems I'm dealing with. Do you think it is worth it's money ? "
Those videos are aimed at hobbyists.
The technical challenges they are finding are in a completely different (amateur) league to those that the serious real world car companies are dealing with.