Automatic anchorperson detection from an MPEG coded TV program

In this paper we propose an unsupervised anchorperson detection algorithm from an MPEG coded TV program recorded for many hours. In order to extract news topic presentation shots, we employed several visual features such as motion, face, caption, and clothing on an MPEG DC image domain. In the experiment, it has been shown that news topics were successfully extracted from a recorded TV program for 24 hours.