Automatic extraction of story units from TV news

In this paper we propose a novel method for automatic identification of semantically consistent story units from TV news programs. The method includes a temporal segmentation procedure of the video based on visual cues, as well as a graph-driven textual analysis technique of the subtitles documents. The experimental results, obtained on a dataset of 50 videos selected from one week video archive of France Television demonstrate the pertinence of the proposed approach.