Camera Motion Detection in the Rough Indexing Paradigm

Abstract

This paper presents our camera motion detection method (pan, tilt and zoom) for TRECVID 2005. As input data, we only extract P-Frame motion compensation vectors directly from the MPEG compressed stream and we so achieve a performance of 3-4 times faster than real time. Our method is based on global camera motion estimation and a likelihood based significance test of the camera parameters. The best run (RI-3) on the TRECVID 2005 test set provides 0.912 for mean precision and 0.737 for mean recall. 1