Essentially, at each time step, the algorithm senses its environment and checks how much it differs from the previous time step, and figures out if it saw any new features to add to the map and how much the correlated features between the two time steps moved, to infer egomotion. This doesn't have to be necessarily with cameras, it can also be done with laser rangefinders and other relatively accurate sensors.

Monocular SLAM (MonoSLAM, also the name of a well known paper) is SLAM done with a single camera, which makes the problem harder than with two cameras. With two cameras affixed to a rigid frame and known characteristics, it's possible to determine the 3D position of any given feature that is seen by both cameras at the same time. With a single camera, however, it's trickier because only the angle of a given feature can be determined, not its 3D position, so an optimization step has to be done to determine what the likeliest solution to the problem is.

There's also more to read on the relevant Wikipedia article, at http://en.wikipedia.org/wiki/Simultaneous_localization_and_m...

I'd like to see the video and learn more.