Personally, I am constantly scanning for people near the edge of the road and anticipating their movements. I bet most of us have had the experience where someone stepped out unexpectedly but you already took action because you could tell they were not paying attention. Good luck programming that.
Yeah there are a lot of subtleties that are hard to program; that's why it must be extremely effective on things that _can_ be programmed such as quick breaking, quick steering; prefer lanes that are farther away from pedestrians near-by, beep when entering a non-illuminated zone and hundreds of other details.
Then it shouldn't be deployed. I had several occasions where I overlooked somebody (luckily without consequences) but never one where I falsely registered someone or something to warrant to brake.
I don't think it will work. Instead the failing effort will be cover to add more and more rules in an attempt to hide the obvious while blaming the victim(s) for breaking the rules.
I actually think this is a fairly trainable attribute. Given that Tesla does this kind of detection for objects and classification and records many miles of human drivers - it wouldn't be surprising that this may be a learned trait.
