"The only exception is when the lines do not exist, or are not visible, in which case the driver will have to take the old-fashioned route and drive with both hands."
That sounds like a severely limited AI.
I wonder if some of these half solutions won't actually be worse for safety. Is someone a great deal more likely to completely shift attention from the road if the car is driving? Maybe even by dozing off? When that happens, how long will it take to regain sufficient focus if the white lines disappear on a stretch of road? How robust is the AI for a car that could be fooled by deceptively-painted lines?
Google's car is much more sophisticated. Given what Google has demonstrated, I'd rather we skip any intervening generations of half-self-driving cars. These pseudo driving cars seem just as likely to give the whole AI vehicle concept a black eye and set the industry back 10 years.
I was just thinking the same thing -- "this will never work in Boston. We fear and hate painted lane markers."
But it turns out that something like 90% of miles traveled are on highways.[1] So what if you just limit it to limited-access highways recognized by the GPS? That covers a huge chunk of miles traveled, and you have actual vehicles out there a lot earlier, which will teach you a lot. And highways kill (I think) tens of thousands of people a year, so getting more reliable drivers out there can't come any too soon.
I'm also thinking that the AI will have to be pretty damn robust to go into production, whatever this article says. Whether or not there are painted lines, it will have to share the road with other cars, so it will have to be pretty smart in order to do anything at all.
No matter how limited the technology, it's good to have automotive companies clearing some of the regulatory hurdles, working on hardware integration, introducing this to consumers, etc.
The principle road tracking was to look for any sets of lines that moved in parallel. The primitive hardware was a series of horizontal line scan cameras aimed at increasing distances in front of the car and a simple sideways shift+correlation to decide which direction.
This way that tracked not just white lines, but tire tracks, kerbs, road wear, roadside barriers etc. The main problem we had was on a clean unmarked newly laid road surface that was perfectly uniform.
There is a highway I commute on daily which has been undergoing roadworks for the last 3years and has so many painted out lines, repairs, temporary markings, filled in trenches etc that in the rain I have no idea where the lanes are.
Made me think of those sections of freeway (I mostly notice this in SoCal, not sure if anyone is familiar w this) where there are grooves or gradations that sort of slowly drift off of the lane dividers on a curve. People often start to follow the false lanes at first. I imagine such a system would have trouble with that.
That sounds like a severely limited AI.
I wonder if some of these half solutions won't actually be worse for safety. Is someone a great deal more likely to completely shift attention from the road if the car is driving? Maybe even by dozing off? When that happens, how long will it take to regain sufficient focus if the white lines disappear on a stretch of road? How robust is the AI for a car that could be fooled by deceptively-painted lines?
Google's car is much more sophisticated. Given what Google has demonstrated, I'd rather we skip any intervening generations of half-self-driving cars. These pseudo driving cars seem just as likely to give the whole AI vehicle concept a black eye and set the industry back 10 years.