its done using software such as boujou or pftrack, you run the video through it and it calculates where it thinks objects are in the scene based on how the images move, it can then make a guess at where the camera would have to be, it's very clever but incredibly slow going.