A Tool for News Video Scene Indexing

Indexing of audiovisual sequences is a difficult task, and it is often necessary to build ad-hoc techniques according to the considered audiovisual class. In this work we present an interactive approach to news programs scene indexing and classification. The entire process is fully automated and the user is allowed to manually correct intermediate and final results. Basically the news video is first segmented in shots. Then, each shot is associated to a codebook obtained by a vector quantization process. Shots in which the anchorperson appears can be easily separated as their associated codebook exhibits peculiar characteristics. Sequences of consecutive shots can be identified forming scenes classified as "Outdoor" or "Indoor" following a certain semantics. This process is achieved by establishing correlation among shots and using the statistical information regarding the temporal distribution of anchorperson shots with respect to all the other shows of a news programme.