I guess the only half-way useful approach would be to use the camera to detect objects and to analyze their relative movement compared to the N900.
In principle it's possible but it's quite tricky, especially if you have no idea in advance what these objects look like and in different lighting situations.
I did something similar when I was a student. But it wasn't that complex, because we at least had a general idea of the objects we were looking for in advance. One of the mayor problems was the CPU power needed to analyze moving images for objects in real time. If there aren't much more efficient algorithms than we used back then even an overclocked N900 will be far from having enough CPU power.