SLAM with one camera

edit

It's possible, but a bit more complicated than using two cameras. You are essentially using parallax in time (the camera position at one instant, and then the camera position at the next instant) to do feature recognition and localization.

Two resources for this come to mind:

1. SLAM with a Single Camera, Davison 2002. I think that he has a bit more information on his own personal website.
2. MPRT's kinect-3d-slam. I believe that this is also using the point cloud capabilities of the Kinect in addition to the vision data. ROS source code is available on their website, though, so it might be a good place to start work.