Hacker News Reader

Comment by neogodless | Hacker News Reader

I don't know how much is from knowing there was a truck in the road, but I noticed in my first viewing how something was blocking oncoming headlights.

If you're going 75mph, that should really get your spidey-sense tingling. At least let off the accelerator/cancel cruise control. Be prepared to brake. Once you're closer and lined up, it's harder to tell there's anything going on because you're no longer approaching a bend with headlights on the other side of the truck, so by then it's too late.

But this is, again, all hindsight talking. In the moment, it's much harder to assess how a "typical human" might have reacted, if at all.

View on HN

Had the same feeling, incoming headlights are not visible = obstruction. pity the car has nothing that could be called intelligence.

I’ve spent a lot of time staring at the road thinking how I’d write a program to do it. Main thing I’d do differently is provide visual feedback of confidence. Having the autopilot disengage when it gets confused isn’t good enough, I want to know when the system is losing confidence /before/ it disengages, and I need it to let me know that it’s not 100% sure the next thousand feet of road is actually clear. Tesla seems to drive unperturbed until its classifier recognizes an obstacle.

As someone who is a decent-ish programmer that's recently worked on recognition from videos, I would be totally lost trying to figure out how to handle a situation like this. As a human person (I swear!), I could see in the video that the headlights were being obstructed by something but, as a human programmer, I'm completely at a loss for how I would even determine if the headlights are blocked by something up close vs. far away using just the camera's view. Headlights get blocked by things all the time. How in the hell would you figure out that that's what's going on just by the visual?

In other words, it's a good thing I don't make this stuff. My car would be phantom braking every 10 seconds, especially at night.

Yeah, I've wondered about this too. If cars in the adjacent lanes slightly ahead of me braked all of sudden, I would brake too even if the road ahead of me is completely clear because I would think there is something going on. I wonder know how this type of thing gets programmed in to self-driving vehicles.

Spoiler alert: It doesn't. It's almost completely based on the what the vehicle in front of you is doing.

I would definitely want night vision infrared, but my main strategy would be making sure the texture of the road surface is flowing toward me at the expected speed. I imagine I could use something like the motion vectors in mpeg compression. I would exploit the fact that the road 1000 feet away is coming at me slower than the road 100 feet away: if there is a visual texture in the middle of my field of view (where the horizon should be: road approaching relatively slowly), and it’s coming at me faster than 1000-feet-away asphalt ought to be, I assume the road is obstructed and slow down until I get a clear view again.

Granted, I am a lowly web programmer with a hobbyist interest in compression, but I have thought about getting the comma.ai kit and learning how it works, trying to implement my own world modeling, but, sounds like a full time job :)

EDIT: as for the headlights, I don’t want to act like I could whip this up in a weekend or anything, but having objects flicker in and out of existence would at least subtract a few points from my confidence score.

That's an interesting way to go about it but would the texture of the road surface be consistent enough and reliable enough to make those calculations? This is the kind of stuff I'm talking about. I definitely do not have the confidence to do that. It's why machine learning is so scary to me. We're basically saying "I can't figure out all the possible permutations of this so let's throw them at a computer and hope it can do a better job".

Can you tell me more about what kind of road (in)consistency you’re describing? It’s fun for me to think of how to improve this strategy.

I am similarly opposed/nervous around machine learning, seems impossible to debug weird decisions the black box makes, I would prefer a traditional, deterministic computer vision approach.

Basically I want to segment incoming images into a grid and do a cosine similarity score between each cell and its neighbors one frame ago, to detect which direction and how quickly all the textures are moving. So I think it should be robust to whatever the road texture is, provided its not pitch black like wet asphalt.

> consistent enough

you want the opposite, if it's consistent, then it's more difficult to gauge it's motion just by finding similar zones. think on mouse on glass or metal, it's too uniform for the sensor to detect motion.

anyway: paper https://dspace.mit.edu/bitstream/1721.1/6298/2/AIM-476.pdf

lecture: https://www.youtube.com/watch?v=0BVZDyRrYtQ

usually circular motion is used, but linear works just as fine

The biggest change I’ve made over 35 years of driving is slowing more and sooner when things are unclear. At 75, just taking your foot off the accelerator will shed 10-20% of stopping distance in a few seconds. Waiting until you are sure it is bad news is the difference between defensive driving and avoidable crashes.

>At 75, just taking your foot off the accelerator will shed 10-20% of stopping distance in a few seconds.

Even more in a Tesla as the regen braking will slow you down a lot. I drive almost exclusively with 1 pedal now because of it and, while I don't know that that actually is safer, it feels safer because I'm slowing down waaaay earlier than I normally would in an ICE-mobile.

It's not going to be reliable with just cameras.

You can use where the headlights disappear and reappear to draw an arc of occluded space. You should be able to measure how quickly that arc changes to estimate the distance

If you're 1000 feet from an object and move a foot closer, it barely blocks anything else from view. If you're 5 feet away, moving a foot closer blocks a lot more of your view.

The problem is that relies on nearly constant backlighting on the object.

The real solution here is LIDAR, which makes this trivial. Tesla won't use it though, so they have stupid problems like trying to hack a way to determine how far away dark objects are.

Exactly. I was just going off of Musk's statements that they don't want to use things like LIDAR since people don't have LIDAR but can make these types of assessments.

Even if you could make an arc of occlusion (sounds like a D&D spell), there's way too much variability in headlight sizes/designs to accurately determine distance in the dark in the situation this video presents. Again, I am admitting a ton of ignorance when making that statement but that seems nearly impossible to me with just cameras.

man, until you said that, it wasn't something i had noticed, and I was staring straight at it with full attention. I can't imagine someone driving for hours picking that up. After all it could an obstruction in the median.

You have the advantage of knowing before hand something terrible is about to go down. The driver was probably driving for several hours at night and has his/her spidey-senses a bit fatigued.

This is a dark truck with no lights in a high-speed road at night, I would probably have crashed too. Yes, human eyes have better night vision, but people have different visual acuity, attention span, and reflexes. Also, headlights from other cars can blind you momentarily.

Still, I don't understand why Tesla is so opposed to LIDAR. The more information the autopilot has, the better decisions it can make.

Lidar is prohibitively expensive. There aren't a lot of manufacturers. Velodyne is pretty much the only game in town, they primarily sell to defense industry. GM via Cruise bought a startup that promised solid state lidar. It ended up being a complete waste of time.

Reply on news.ycombinator.com