For some time now they've been able to detect temporary road signs from amongst advertising and other general signage. i.e., changes in speed, maintenance warnings, etc.

If a human can see the child on the bike behind bushes, the car can too. Plus it can do it from multiple cameras at once. And calculate the speed and trajectory. And quickly work out the optimal course of action.

I really think you're overestimating the ability of humans (who get drunk or distracted, routinely speed, etc) and underestimating the progress of self-driving cars.

I'm not overestimating humans, I never said they couldn't fail. What I base my comments on, is people I know that study AI research (mostly graduate students and some doing their doctorate). All of them say that even state of the art AI for a specific purpose, let's say something as simple as face detection (not recognition, just "this is a face") fails for a lot of inputs.

That is not to say some people that are drunk or tired and make the same mistakes as an AI in a driving scenario - the difference is that a computer will always be confused by a scenario that it cannot understand (and in worst case misintreped it), whereas humans have a much easier timer making sense of such situations.

Until we have a strong AI with sensory capabilities akin to our own, I doubt we'll see truly selfdriving cars that require no human intervention. I'd love to be proved wrong, but the research is having a hard time breaking the last 1% needed.