22 comments

[ 3.0 ms ] story [ 54.8 ms ] thread
(comment deleted)
Sounds very interesting. After reading the press release though, I can't find any references to MEMS or why/how this is better than existing LIDAR sensors. Any other resources OP?
Page 2, in diagram. No motivation though.
I believe it is smaller and cheaper, not necessarily better.
Cheaper LIDAR is good, but my money is on an array of super cheap cameras and better image recognition to get the additional cost of "self driving" down low. Also, cameras are lower profile than LIDAR (looks matter in cars...)
Doesn't help you when visibility is poor, though. Of course, humans are no better in those conditions, but a big selling point of autonomous cars is that they can (in theory) be perfectly safe even in dense fog, snow, etc.
Can't infrared cameras see through fog?
LIDAR gives you single sensor depth perception though without having to go through computationally expensive stereo vision mapping. There's definitely some benefit to that if it can improve your latency (or "framerate") such that you have extra time to swerve, slow down, etc.
Many stereo vision systems have an ASIC or FPGA that automatically calculate the depth information from the two images, like this guy:

http://carnegierobotics.com/multisense-s7

Yeah, and it's only 15 fps so that means at least 60ms between decisions. 80 mph is 117 fps so 60 ms represents 7 feet. If you ran at 30 fps then that's only 3.5 feet, at 60 fps it's 1.75 feet.

I'll grant you that saving 5.25 feet at 80 mph is definitely going to make a difference very rarely, but it will happen to some degree.

Also there's no pricing so I suspect it's super expensive, which is not at all what the parent was suggesting "but my money is on an array of super cheap cameras and better image recognition"

I suspect that LIDAR is also going to be fairly immune to glare where as vision systems are probably susceptible. Having a big depth hole open up in your mapping/planning system because of glare would likely bring the car to a screeching halt (literally) which isn't going to be good for the other cars around it.

Lidar also, at least potentially, allows you to focus sampling on areas where you have less confidence in what you're seeing.
I think if I was going to ride in a self-driving car I'd want it designed like a hard realtime system where it is guaranteed to sample the road at X rate or whatever. Most of the time it won't matter at all and doing something adaptive would be fine.

But when the cameras go out completely (or get blinded or whatever) then you could end up with a really degenerate case where the software can't keep up with the car's motion and it freaks out and crashes (literally or figuratively).

Prior to suggesting that this couldn't happen, or would be extremely unlikely to happen, please remember the results of the Toyota "unintended acceleration" investigation: https://news.ycombinator.com/item?id=9643204

If you only meant that the passive system described by the person you were responding too had too high a latency then I think we are in agreement.

However, regarding binocular versus LIDAR in general, I would say that while a greeter frequency is always better 15 Hz should be fine. And in fact the awesome $70,000 LIDAR used by Google's self driving car only runs at 15 Hz.

http://velodynelidar.com/lidar/products/brochure/HDL-64E%20D...

Stereo vision does little at long range, unless you have very wide camera spacing. Human stereo vision doesn't provide much depth info beyond 20m or so, and that's under good lighting conditions. LIDAR gives you equally accurate range data for anything it senses, regardless of lighting.
It's really useful to have multiple sensing mechanisms that have independent error behavior. I suspect when self driving cars are out there in mass, they'll use a combination of cheap cameras, lidar, etc for reliability and redundancy.
Generally LIDARs work by rotating a mirror that the laser bounces off of giving you a 2D plane or cone where you can detect things. Super expensive LIDARs like the one on the Google car use a number of cones with different angles to approximate a 3D system and cost tens of thousands of dollars. Other systems like the ones you see in the DARPA challenge might rotate the plane of the LIDAR through a full circle, but that takes a while.

Spinning mirrors are also somewhat delicate which can be a problem for automotive applications.

Exactly. That expensive rotating Velodyne thing is a research tool, not a production auto part. Production cars need better, preferably with no big moving parts. The Pioneer MEMS device qualifies (the moving parts are at IC scale). The new Kinect is a no-moving-parts 3D LIDAR. The Advanced Scientific Concepts flash LIDAR is also a good 3D LIDAR, usable outdoors in daylight. It costs about $100K because it's hand-made by physicists in Santa Barbara for DOD and Space-X, but the technology isn't inherently expensive.

The Pioneer press release doesn't say much. The usual specs are X and Y pixels, field of view, and range, none of which they give.

(comment deleted)
I suppose they have a way to cheaply mass-produce lidar scanners. Not an expert on the topic, but what are the main advantages compared to cheaper Kinect style sensors? I suppose they are much more accurate, but would a difference in the order of mm matter for self-driving cars?
Neat, it's too bad they seem to be tying it into some awful software and mapping system. Just sell the part that you're good at and let other people do the software!