Since you're one of the few people here doing more than speculating, could you expand on this a bit? Is the LIDAR resolution a hard physical limit, and if so what's the origin? What makes identifying the debris with the camera so hard, just the standard object recognition problem in computer vision?

Certainly. LIDAR resolution is related to a few things:

- How many beams you're putting out and at what angles

- How fast you're spinning and how fast you can get the data back.

- How many points of data you can get through your bus

If the Google Car is using the big Velodyne (http://velodynelidar.com/lidar/hdlproducts/hdl64e.aspx), it puts out 1.3 Mpt/second over a 26.8 degree vertical field of view. That makes it put out ~135 pts per vertical and horizontal degree. At a range of 50 meters, with angular resolution of 0.09 degrees and 0.4 degrees vertical resolution, you're looking at vertical spacing of 34.9 cm/pt and a horizontal spacing of 7.8 cm/pt. You'll need multiple "hits" to make an obstacle, and they have to reflect well enough to be counted. A common thing to do is reject any obstacle below a certain height - since it could just be noise.

Debris with cameras is tough because you can't differentiate between "dry spot on pavement" and "someone's dropped a sheet of drywall".

But you are filtering the LIDAR data (source: I have I Velodyne LIDAR in my office). Yes, one stray hit will be ignored. But, you aren't getting one stray hit. You get 20 hits on this revolution, 6 on the next, 21 on the next, and so on, and then use various algorithms to determine if you are seeing objects or noise. Recall that besides the inaccuracy of the spinning lidar (it does not hit exactly the same spot even when sitting still and hitting an immobile object), the car is moving. Any decent Bayesian filter derives a heck of a lot of information from these changes. I'm not talking about the physical change, which is also important (change of position = change in angle = different reflection point). I'm talking about how your process model generates information via hidden variables - from position over time we derive velocity, and the correlation between velocity and position increases our accuracy. In the context of the LIDAR, the filter can detect that there is a small clump of points moving towards the car at 70 mph (from it's point of view, of course from our POV it is the car approaching the debris at 70 mph). Reflections that are true noise will not be clumped, and will not consistently have a velocity that matches the car's.

With all that said, I've just played with the LIDAR, I can't give you lower detection limits. But with proper algorithms it is better than the static computation suggests.

Of course. Filters do wonders - but they're still not a panacea.

You also get the covariance across that filter. Which means you still need to decide how much to cut, whether or not to reject as rain, dust or a reflection, whether to trust your reflections under a certain reflectivity value. Data helps with this, but, as always, there's a ton of corner cases.

Thanks for all the LIDAR info!

A human seems to be able to differentiate pretty reliably (although not perfectly) between dry spots and drywall. And at 50 meters, I expect that binocular vision doesn't do much.

I wonder if people can use parallax (i.e. road being revealed differently) for a 2 inch object like drywall at those kinds of distances and speeds. Parallax is uses in Google's latest software lens blur, so I'm sure it's been in robot cars since the beginning.

Less that perceive the depth (2 inches at 50 meters is tough, especially from above) but that we're really good at pattern matching (vague square thing in the road? likely not road).

You can think of parallax as stereo vision repeated over and over again with different cameras. Cameras/depth sensing has a hard time at range, since the resolution for depth decreases rapidly towards subpixel.

Could you use some kind of LIDAR stereo configuration to tell the difference between a dry spot and a solid object?

Well, LIDAR doesn't have that differentiation problem - it has a resolution problem for small (sub half meter) things at range.

Stereo LIDAR would just add more points - you can correlate the two, but you'll still have resolution issues.

Well there are "Flash Lidar" sensors, that are effectively instantaneous depth cameras firing a lidar beam for every pixel - but a) they're hugely expensive, and b) they don't seem to have taken off (possibly because academia can't afford them to do the interesting research that'll help them take off)

There are some cheaper ones now - the K4W2 is a flash LIDAR.

But range is always difficult. I think PMD is in use in BMWs. They purport a 90 meter(!) range, but I've never played with one.