These algorithms can allow you to have a self-driving car with only cameras. But, there would be a lot of problems if you tried to make a camera-only system for consumer vehicle navigation. Vision systems need distinct "features" in images to find and track across frames to allow you to compute distance, speed, etc. If you don't have many features, pure vision approaches won't work. Nighttime operation is a big problem, as is driving on relatively smooth, featureless terrain.

The basic downside is that standard consumer cameras are passive devices. That's why Google uses LIDAR- it's an "active" technology that creates its own features. And driving is an application where the usual computer vision "it works most of the time" is just not good enough. Time of flight cameras are interesting sensors that combines active with passive technology. As this technology matures it might allow for self-driving cars without LIDAR.

Thanks for that explanation. I have been thinking of experimenting with automated driving with cameras and got encouraged by things like these: https://www.youtube.com/watch?v=dcm9NpMNi68 . But yeah, I can understand how ideal conditions are very different from the real world.

Yes. Machine vision has advanced a lot in recent years and it might just be possible. There is at least one startup trying to make self-driving cars with just machine vision.

Computing the differences between several cameras can be a judge of distance, but you can also see how much the object moves as the car moves, and get an estimate based on normal machine vision (how big objects like that normally are, objects nearby it, where it's shadow is, etc.)

Do you happen to remember the name of this startup?

I have been thinking of experimenting with automated driving. Have started reading up on computer vision and found some encouragement from things like these: https://www.youtube.com/watch?v=dcm9NpMNi68

In the future maybe. Computing the depth maps in this simple case is not very costly because it only requires a relatively sparse amount of key points. To recreate the full scene geometry around a car while moving at 100kmph is a lot harder. Not only do you need to use a massive amount of key points but you need to produce a depth map in millisecond time frames.