the problem you mentioned, that people had to move back and forth to calibrate is already the key point of your problem.
This is due to the fact that to be able to recognize people in a scene, openNI needs some initial movement of the subjects.
As soon as point clouds are classified as persons, the calibration can start.
since you only started the recording AFTER the people were recognized, (I assume) they don't move, so are not recognized...
if you had started recording from the beginning, this would have worked from the recorded files the same way as from the live video feed.
i'm sorry i have no idea of how to help you with your problem except that you need to recapture those videos...