Re: Self Driving Cars: The Future is now and by now I mean s
True; if that sort of infrastructure were put in place, it would make things much easier. The Urban Challenge required designing a car that did not rely on such things, however, but only used existing infrastructure.
If all you've got to work with is the stuff visually available to a human, then you need:
1) Accurate feature tracking to estimate movement
2) Foreground object tracking
3) Foreground object recognition
4) Foreground object motion prediction
5) Optical character recognition on street signs
6) Motion Planning
Those all have known approaches, but they aren't reliable enough to trust a car to yet. Especially not in combination. Some RFID would help particularly with the OCR and object recognition tasks.
Christopher said:
Lindley said:
The big problem with GPS navigation being the primary key point is that you can accumulate fairly large position errors when you go under a bridge or in a tunnel. You can compensate to some extent for that using inertial motion sensors, but they compound errors fast on their own. If you integrate visual navigation as well by keying off feature points seen by a camera, you might have a chance.
GPS isn't the only way to go. As long as you're on known roads that are clearly marked in a way the car can detect (say, RFID tags at every intersection/exit/etc.), then it can rely on internal maps and local sensor inputs. I mean, a person driving a car doesn't need to know their latitude and longitude to the arcsecond, so why should a computer? When dealing with a lot of individual units in a system, it makes more sense to have them rely on local rules than global control systems.
True; if that sort of infrastructure were put in place, it would make things much easier. The Urban Challenge required designing a car that did not rely on such things, however, but only used existing infrastructure.
If all you've got to work with is the stuff visually available to a human, then you need:
1) Accurate feature tracking to estimate movement
2) Foreground object tracking
3) Foreground object recognition
4) Foreground object motion prediction
5) Optical character recognition on street signs
6) Motion Planning
Those all have known approaches, but they aren't reliable enough to trust a car to yet. Especially not in combination. Some RFID would help particularly with the OCR and object recognition tasks.