Hacker News new | ask | show | jobs
by stevoo 2642 days ago
This not news. This has existed from the first iteration of Tesla Autopilots. That is why there where crashes with stationery vehicles for Tesla. This is shortcoming for using only cameras or any vehicle that has adaptive cruise control. The problem is that the front camera cannot distinguish from the road and a stationery vehicle. For that, they are the same and they ignore stationary obstacles. If they where taken into consideration, then they would not work at all.

This is where Lidar comes in but at its current price, or at price that Tesla was designing its system it was too expensive.

I am not entirely sure that pure cameras will ever be enough

3 comments

The Subaru EyeSight system uses cameras. It appears to do pretty well for adaptive cruise control and front crash prevention.

https://www.iihs.org/iihs/sr/statusreport/article/53/3/2

Not sure why you are downvoted. My Subaru does a pretty good job of detecting both moving and stationary objects, using only stereo cameras. It does have its limitations, and I wouldn't trust it to be a level 4 or 5 auto-auto without additional sensors. But it sure blows away Tesla when it comes to little things like detecting parked fire engines and highway barriers.
Seems like eyesight does a pretty good job at 35 mph, not so much at highway speeds.

Have you had any experience with eyesight helping at highway speeds?

Guess that depends on what you mean by "helping".

What you describe is specifically related to emergency braking. None of these systems claim to prevent collisions at high speeds through emergency braking. But they will significantly reduce the impact. No, I haven't tested that personally, and I don't plan to, so no actual "experience" there.

However, there are other aspects of EyeSight I do have experience with at highway speeds.

- Collision alert. It's an extra set of eyes to keep an eye out for trouble. When I'm checking over my shoulder before a lane change, I can take a little more time and look more carefully, knowing that it is watching the road ahead. It has alerted me for things like a car slowing suddenly or squeezing in front of me.

- Adaptive cruise. This helps by reducing load on me while driving and maintaining a safe following distance. So I can focus less on the car ahead and pay more attention all around.

Tip for users of EyeSight and other systems: Always remember that like every other system out there it's not perfect and has limitations. Don't get complacent. After 5+ years of driving with it, I'll still keep my foot ready to brake, even in situations I'm confident it can handle. Occasionally I'll give in to the urge to brake only to find that it has started braking by the time my foot hits the pedal. But I'd rather react too soon than wait too long.

This is something I don't understand. If you have more than 1 camera, you can run an edge detection algorithm on the images from each camera and use simple trig to build up a 3d model of what is in front of the car. This can easily distinguish between a shadow or road kill, vs a large object in the way. Is there a technical reason they don't do this (such as too much processing power required)?
You are trivialising the problem. Getting reliable data from stereo images is still far from solved, especially in realtime. (Also, the best currently available techniques are likely too computationally expensive to run on the compute that Tesla has put in the car).
You can clearly see here (https://www.youtube.com/watch?v=ROHW3sm2j0g) that Autopilot is detecting moving, stationary, stopped and oncoming vehicles.

There could be situations where this does not work accurately and that's when problems occur. If you watch closely you can even see it sometimes in this video.

Works 95% of the time is not good enough.
Couldn't a secondary camera help more? Matching the images is not trivial and may take a while, but you could extract spatial information. Maybe a laser or a flash projector works better, but I imagine these systems to be susceptible to disturbances like reflections, especially when multiple cars are driving around.

Probably a long way still until we have real self-driving cars...