Slashdot Mirror
Pioneer Looks To Laserdisc Tech For Low-Cost LIDAR
itwbennett writes: Pioneer is developing a 3D LIDAR (light detection and ranging) sensor for use in autonomous vehicles that could be a fraction of the cost of current systems (the company envisions a price point under $83). Key to this is technology related to optical pickups once used in laserdisc players, which Pioneer made for 30 years. From the ITWorld story: "The system would detect objects dozens of meters ahead, measure their distance and width and identify them based on their shape. Pioneer, which makes GPS navigation systems, is working on getting the LIDAR to automatically produce high-precision digital maps while using a minimum of data compared to the amount used for standard maps for car navigation."
ARGHH!!!!
https://www.sparkfun.com/products/13680
The operational description is pretty cool too.
GPS location information has revolutionized mapping. Google and Apple get high quality position data for cars traveling. Any non-mobile platform system have many barriers to entry (cars traveling around taking photos), verifying road information. Google gets to validate incredibly accurate maps under the guise of providing traffic, location services and so on the the mobile phone users.
LIDAR, particularly those paired with accurate color mapping information for those point clouds are going to be creating high accuracy, full color 3d meshes of the pretty much anywhere that cars go near. The small snippet of video just after https://www.youtube.com/watch?... (from the previous slashdot article about autonomous golf carts) shows some of the power of LIDAR. Just after the car flyby it shows the point map being updated by every pass through an area. Incrementally correcting the 3D view of the area. The example is just a few passes by the same car.
Imagine 1000s or 100,000s of automonous vehicles with LIDAR pushing data to a vendor for "navigation experience", but still building a 100% model of everything line of site to the LIDAR. And then imagine 100s of drones flying overhead, doing virtually the same thing.
Throw in some AR like Magic Leap being able to deal with that 3d Mesh. Pretty amazing things that our kids will grow up, pretty scary for everyone else..
Unless they just want the small/low speed vehicle market, their requirements are grossly lacking. For a road-going autonomous vehicle, "dozens of meters" is useless at any real speed. Try ~100 meters as a start and go from there. Otherwise you'll always be outrunning your perception and unable to see anything in time to stop.
Gay Man Walking!
Sleep your way to a whiter smile...date a dentist!
The cynic in me says that Pioneer is sitting on warehouses full of 780nm IR laser diodes (used in LaserDisc and Compact Disc players).
Calling them "LIDAR components" will allow Pioneer to mark them up at year end.
At highway speed, 70 MPH, a car needs 75 meters (dozens of meters) to come to a dead stop. Human reaction time requires further distance, but that doesn't apply to fully autonomous vehicles. It only partially applies to a technology-assisted safety system which augments the human driver.
I've collected Laserdiscs and players for several years, though haven't really been into it much lately. I have more than 500 movies and at least four working players...
I can now build an autonomous vehicle!
WOOHOO!
Do not look into laser with remaining eye.
Sure, if you assume no processing time is required to characterize what the sensor sees, and want to slam on the brakes at a minimum of 0.66g's of deceleration the moment anything enters the edge of your perception. That doesn't seem like a viable human-carrying vehicle to me.
I can now build an autonomous vehicle! WOOHOO! http://bit.ly/1EoywpE
http://bit.ly/1MSFGuI Horror movie ...I can see it
really? that is what they are famous for?
I have to google it...ah for cars:
http://www.pioneerelectronics....
4wdloop
Oh god, please don't let this happen. My police laser detector will be completely worthless.
I find the original article and the one at slashdot to be spectacularly void of information.
I remember the laser disk and the one kid who owned one. The rest of had VHS and thousands of titles. 30 years? Really?
Think about the video game implications! All that 3D data! It was everyone's dream to play your FPS in your little home town... imagine a whole virtual world. World organized virtual battles... The Chinese are invading Canada, quick get online and do your part citizen! :)
> a car needs 75 meters (dozens of meters) to come to a dead stop
Thank you for translating that into Imperial units for my non-metric brain! :)
Is lidar also used for the digital recreation of planet's surfaces?
> Sure, if you assume no processing time is required
Let's give the system time to do a million processing operations. The GPU and CPU should be running at 1 Ghz or better, so we need 0.001 seconds to run a million operations.
70 MPH is 103 feet per second, or 1236 inches per second. So in the .0001 seconds required to do a million processor ops the car will travel 1.2 inches. You figure it could do more extensive processing, 10 million operations? Okay, that's 1 foot of travel.
Computer processors are really, really fast. Hard disks not so fast, but this system isn't reading the data from a hard drive.
> slam on the brakes ... the moment anything enters the edge of your perception.
?!?!? I'm not sure how that comment has anything at all to do with what is being discussed. The ONLY time "slam on the brakes" would ever make sense would be if there was a large stationary object directly ahead. Most cases requiring action are for moving objects on a collision course, such as someone on a cross street who is on course to T-bone you at the upcoming intersection. Since the other car is moving, you can either tap the brakes and let them pass in front, or speed up a little and go through first. Or turn. Only a light tap on the brakes is needed from 100 feet away in order to give the other car the extra 1 second they need to pass in front of you rather than hit you.
The way it works, for humans or computers, is that you measure the distance to the object and it's vector - direction and speed. The hardware does this measurement rather faster human wetware does - it's a time differential thing, so it can be done directly in silicon, no code required if you want it very fast. Figure about 1,000 microseconds, or 1 ms. This measurement is passed to the processor, which does a bit of math to determine if the object is on a collision course with the car. Remember CPUs are really fast at math. That's the 1ms we discussed above.
So we have about 1 ms to detect vector, and 1 ms to determine whether or not it's a collision course. The car has traveled 2.4 inches during this time. Assuming it is a collision course, we continue. We already have these two pieces of information:
1. some is coming from the left
2. it is scheduled to collide with the left side of the vehicle, 120 feet from now
Given those two inputs, the car now needs to decide between three options:
Slow a bit, allowing the other car to pass in front
Speed up a bit, getting through the intersection before the other car
Turn*
A million operations, or 1ms, is plenty to choose between the three.
* "Turn" is mainly useful if you assume the other car is TRYING to hit you, so rather than continuing to go straight they'll turn and you don't know where they'll end up. The one thing you DO know about where a speeding car will be three seconds from now is that it won't be in the same place it is now. So steer toward where it IS and you can be pretty sure it won't be THERE any more.
I didn't say it was easy. I said processing is fast.
Yes, such systems need training. You don't, however, need to wait for it to retrain every time it spots an object in your path. You don't even need to
instantly recognize that it's a car. Once you read that it's a) dead ahead and b) not moving, you can compute that you should start slowing down.