You get to use worse hardware, but you need several order magnitude better software.
I expect that, after the first wave of clumsy LIDARing self-driving cars, all the car companies (Google especially) will be collecting exactly training data from the cars' sensors to build exactly this kind of model. In fact, I wouldn't be surprised if that was what the Google car was really about, in the same way Google Voice is really about collecting speech training data.
The best part of this kind of training data is that it all comes pre-annotated with appropriate reinforcements: even if the image-recognition sensors aren't hooked up to anything, they're coupled to the input stream from the other car sensors and the driver's actions. So you would get training data like
- "saw [image of stopsign], other heuristically-programmed system decided car should stop, driver confirmed stop."
- "saw [image of kitten standing in the road], other heuristically-programmed system decided car should continue, driver overrode and stopped car."
Etc. Aggregating all these reports from many self-driving cars, you could build an excellent image-to-appropriate-reaction classifier.
In a driving situation, the worst-case scenario is everybody dies.
The human eye can instantly recognize the available driving paths, the motorcyclist ahead, and project where people will walk. Software would have to parse out where the open roads are, how far that motorcyclist is and whether he can clear the intersection before the car reaches it, and what that sign on the right-hand side is—using the same information, but it has to parse it first whereas we do that almost instantly. It's a totally different game.
Edit: a better solution would be to observe the behaviour of other drivers; if there is someone driving ahead of you, you can assume that the road between you and them is OK; if there's noone ahead of you, you need to drive slower and be more careful (that's how I drive at night). Once there's a critical mass of cars with cameras, cars could communicate road conditions automatically.
I would be frightened to trust the data coming from a random car in front of me. Inferring road condition from another car behavior sounds reasonable, using data supplied from it, not so much.
However, if the car three cars in front of you just broadcast "I'm doing an emergency stop right now", that's really valuable data. The human in your car won't know anything is wrong for at least a second. The human driver behind you would know about it before the human driver in front of you.
Until some asshat implements a button that sends this message for "lulz", though that would probably constitute a crime under existing law.
A computer-driven car, though, wouldn't (shouldn't) just immediately slam on the brakes because of that signal. It would tighten seat belts and start slowing down, but it also would want to avoid getting rammed by the car behind it. It can make very accurate estimates about its stopping distance and use all of it.
That's just my theory on it.
The simple answer is that we don't know how to do it with stereo vision alone yet. Getting range reliably is hard, and your brain uses lots of tricks to do it.
The second answer is that we need better-than-human performance if this is to take off. So using human-type sensing might not be good enough anyway.
Lots of researchers are pushing on vision-based driving, though.
Humans proved that for current level driving two not particularly high-resolution cameras are sufficient. Seems like pushing in this direction will remove this expensive component?