Dynamic obstacles aren't the end of the world with modern computer vision libraries, even with just a rgb webcam, not using depth sensors like kinect.
2010 OpenCV and shitty rgb camera tracking lanes/road/cars:
[ame=http://www.youtube.com/watch?v=dcm9NpMNi68]Road lane and vehicle tracking OpenCV - YouTube[/ame]
It's crude, but a company like Google or Microsoft could do a lot better. Especially as infrared/other depth sensors get cheaper/better and you can have hundreds of sensors on a car. Some sort of night vision/heat sensor could detect warm blooded obstacles.
Not saying it's easy, just that it's not that far off.
2012
[ame="http://www.youtube.com/watch?v=vpsA4iZ5BdY"]Real-time lane detection and tracking in embedded systems - YouTube[/ame]