|
Dumb question: how do these systems distinguish a traffic light? Here in New England at least, there's quite a variety of lights, many different modes (red, yellow, green, flashing red, flashing yellow, no-turn-left red arrow, no-turn-right red arrow, and different light technologies: old-fashioned style, Fresnel lens, slotted shade. Add to this complexity the weather conditions. Suppose the sun is shining straight at you and you need to squint and shade your eyes just to make out what the light is -- this happens to me frequently -- can the camera see the traffic light and distinguish its color clearly under such conditions? What about when it's raining, misting or drizzling, snowing heavily, etc. and the traffic lights are these fragmented outlines that you, the human, can heuristically distinguish but a machine might not? One last thought: suppose it's right turn on red and first car in line is a self-driving vehicle. Can it really look left and safely determine there's enough time to beat the cross traffic? If it's highly conservative and just waits until green, there could be ten irate motorists behind it and guaranteed to honk and curse. It's exciting technology but there are some very difficult problems to solve. I worry that if these machines can't demonstrate 110% of a human's ability to drive, they simply won't be implemented in many places except some very well defined rigid routes that are free of problematical challenges and variations. |
Well, how do you distinguish / identify a traffic light?
For me it's a combination of knowing the area, knowing what a traffic light looks like, and observing the behaviour of traffic around (mostly in front) of me (if the intersection isn't visible).
For an autonomous vehicle, they'd use the same methods plus they'd have the non-trivial added benefit of colleague robots feeding them updated information.
> Suppose the sun is shining straight at you and you need to squint ...
I'm sure the human eye + sunglasses + visor comes a poor second compared to CCD + mechanical + IR + (etc) can do.
> What about when it's raining, misting or drizzling, snowing heavily, etc. and the traffic lights are these fragmented outlines that you, the human, can heuristically distinguish but a machine might not?
Interesting use of the word heuristic there. If you can determine the heuristics you are using, then an autonomous vehicle can use the same.
> One last thought: suppose it's right turn on red and first car in line is a self-driving vehicle. Can it really look left and safely determine there's enough time to beat the cross traffic? If it's highly conservative and just waits until green, there could be ten irate motorists behind it and guaranteed to honk and curse.
This, and variations, frequently come up in discussions on AV. There's an implicit expectation that no one building these things has considered this problem (this is clearly false). There's two explicit expectations that once a tipping point of AV's are out there, a) there'll be a regulatory push to massively accelerate the adoption close to 100%, and b) inter-vehicle communication means this problem won't arise. In the short term these problems may occur, but to address your question, yes, I'd suggest a computer would be better able to predict if there's enough time to safely turn than most humans.
> I worry that if these machines can't demonstrate 110% of a human's ability ...
Which human would you pick?
Anyway, I wouldn't worry - none of these problems are intractable.