Hacker News Reader

Comment by Dwolb | Hacker News Reader

LIDAR is pretty cool because you don't have to do object recognition on the point cloud to know what you shouldn't be bumping into - since the data comes back "already 3D" you have distance and size information and just need to calculate speed. Another benefit LIDAR has is its high resolution as compared to RADAR.

Where LIDAR typically falls down is in adverse conditions (rain, snow, low sunlight in the horizon). There's a group at my university, MDAR, who is prototyping a tech to better overcome the adverse weather issues through a better laser sensing technique.

Here's an article with a little more technical information [1].

Disclaimer: I've done some part time work on market research for MDAR.

[1] http://articles.sae.org/13899/

View on HN

you don't have to do object recognition on the point cloud to know what you shouldn't be bumping into - since the data comes back "already 3D"

This is true of any pointcloud system though. No system returns density though, so you still have to make assumptions about what is worth running into at what resolution. So if a leaf falls in front of your car, you don't immediately slam on the brakes.

I don't know of any vision systems that do object recognition immediately, in fact I'm not even sure that is possible - or desirable. I guess if you baked segmentation into the firmware it would serve as a fast classifier, but AFAIK nobody is even talking about doing that as you would need a very well trained vision library and classifier to deploy to FPGA.

I make the distinction because we build fairly dense pointclouds with monocular RGB cameras and VSLAM. LIDAR isn't required anymore. LIDAR is faster with lower error tolerances today, but nowhere near as cheap or portable. I suspect monocular SLAM will overtake LIDAR in the next couple of years as it's getting faster and more accurate everyday.

I'm not sure if I understand you correctly. But Mobileye, probably the most prominent company in the field, does object recognition? And they design their own ASICs.

https://www.youtube.com/watch?v=jKfwHsHUdVc

Right, but you don't have to in order to get 3D depth data from a point cloud.

You make really good points all around. The part we're not so sure about is if there's no place for LIDAR when discussing autonomous driving.

As you mentioned we'll have to see where economics takes over as cameras are much cheaper, but less accurate.

I'm curious if your camera-based solution is for outdoor or automotive use. If it is meant for outdoor use, how is performance with oncoming headlights, a low-horizon sun, or in snow?

Ours is all environment, but for mobile AR and not autonomous driving. So our draw distances don't need to be as fast and we have more persistence and more frames per meter than a driving system would have at 60km/h. That said, if we weren't restricted to mobile hardware we could probably get close to matching LIDAR speed and resolution.

The snow thing is interesting actually, but in a different way than you might think. Because we do loop closure it makes matching scenes that are temporally separated (eg. no snow vs snowy or occluded by leaves etc...) very hard.

I agree VSLAM is good and getting better everyday, but I don't necessarily agree it'll displace LiDAR for autonomous vehicles. The first generations will probably have both. Sure, LiDAR is more expensive, but we're talking about avoiding collisions here -- I believe the extra safety is worth it. I don't think portability is not a big issue for cars (though it is a bigger concern for smaller aerial vehicles). LiDAR will still work in the dark, and potentially with much more range than cameras can. Having worked with vehicles using both technologies, I'd feel far better on the LiDAR-equipped one.

100% agree currently. Your point about dark I was totally overlooking, so I think that's the big difference that makes LiDAR better long term until something better comes along.

If LIDAR becomes cheaper though, then what wins? It should take less processing and mean lower latency I would think.

> you don't have to do object recognition on the point cloud

This is true, as long as you're willing to treat everything your LiDAR sees equally. The problem with this is that you're treating, say, a patch of grass (which you might as well drive over), a rock (which you'd rather drive around) and a pedestrian (which you might rather stop for, as a moving obstacle) the same way. Object recognition is still a good idea with LiDAR. And I'm not even getting into noisy returns or "negative" obstacles (i.e. holes).

Disclaimer: has done research in object recognition with LiDAR ;)

Bottom video on this page is my favourite example of this - one 800kg autonomous vehicle carefully avoiding stepping on the grass...

http://www.drtomallen.com/publications--videos.html

What I wonder about for LIDAR is how it will work when you have 20+ Cars using these devices simultaneously, say when driving on a busy road. Do they suffer from not being able to discriminate their own signals from those of the other cars, or is that (statistically) not a concern?

I suspect that a little bit of modulation cleverness goes a long way towards eliminating crosstalk.

Not sure if you look at military applications then when a group of aircraft will fly in formation only 1 of them would have the radar on in active scan mode and the data is usually going to be shared over the data link. With phase shift due to movement and the uncertainty of how many cars are there around you modulation might not be enough to prevent cross talk.

Yes, overload is real, you can never eliminate crosstalk only suppress it, and degenerate cases will probably pop up as the tehcnology scales. But coming to a compromise that allows every car to generate a map of the road should easily be possible. Well, "easily" in the technological sense, perhaps not the political sense.

For military applications it's more useful to have one radar at high range/power than to have n radars at low range/power, while it's just the opposite for civilian lidar applications. So I don't see military radar SOP as evidence that this is inherently difficult.

Do they still do that? If so it's because they don't want to have frequency agility turned on all the time so potential adversaries get less opportunity to figure out the relevant algorithms. The degree by which radars adjust their pulse frequencies is way larger than the doppler shift of anything you're likely to run into traveling in the atmosphere. You will occasionally have a conflict by chance but radars by their very nature have to deal with a certain amount of noise.

With lidars you won't be able to adjust frequency because that's not how lasers work. Well, you can with free electron lasers but nobody is going to be putting those in lidars. But a laser is directional in a way that a laser isn't. You will occasionally get some false returns but real objects persist from scan to scan in a way that noise doesn't. I've worked with a small fleet of lidar equipped robots and lidar cross talk never caused us any problems. Sunlight on the other hand could be a big problem but automotive lidars are all rated for the outdoors.

Ah yes - the ol' "facing west near sunset" bug... I lost a few nights sleep to that one. :-)

This is why I believe LIDAR is a blind alley. We will need to use stereo camera setups and clever bit of firmware (maybe even neural nets) to extract 3d directly from the picture in real time, just like the rest of us monkeys.

With a flash LIDAR, it's not a big problem. If you're taking 100 images a second at 200 meters, you're only receiving for 1.2us every 10ms, for a duty cycle of about 1/10000. Conflicts are possible but unlikely. If you vary the flash timing randomly, the odds of N successive collisions are (1/10000)^N. A frame with a conflict is going to look totally different than the previous and following frames, so it will be obvious when you've been jammed.

What would be stopping people from driving around with jammers, or how about enthusiasts who would mod the LIDARs on their vehicles to achieve a higher frame rate? Will police have detectors on their cars, patrolling for these vehicles that are not following the spec?

The same thing that stops people driving around with GPS and cell-phone jammers. That is, nothing, apart from the threat of prosecution if caught.

This seems to be a common complaint about new technologies; people claim that somehow anti-social, violent or abnormal behaviour will be a problem merely because it is _possible_.

You can blind one of the things with a laser pointer at the right IR frequency. But the device can easily detect this, locate it, and use a camera to take a picture of the jammer. For directional devices, jammers are targets; you get direction, but not range.

The other unintended consequence of not doing any kind of material recognition is that the system will treat a plastic bag or piece of styrofoam the same as it would a large piece of sheet metal. One of those requires evasive maneuvers and the others do not.

If LIDAR could sweep materials with varying frequency and measure the reflection vs. absorption across a band, you could get pretty decent material identification.

There are some conditions that LIDAR handles badly. We discovered that the charcoal fabric used on many office chairs is a very good IR absorber. The SICK LMS won't see those things at point-blank range.

Some LIDAR units give you "first and last", that is, the time to the first reflection and to the last one. For a solid surface, both values will be very similar. For snow, rain, and foliage, they'll differ quite a bit. This is used in aerial surveys, where "first" gives you the tops of trees and "last" gives you the ground. It hasn't usually been available for ground-based LIDAR units, but there's no reason it couldn't be.

Isn't the same true for stereo camera ?

Reply on news.ycombinator.com