Unsupervised story segmentation and indexing of broadcast news video

This paper presents a novel technique for story segmentation of news videos. The visual similarity, silence in the audio and the text in text boxes of a news video are used as parameters to define the story boundaries. Each of these parameters is used to create an index and these three indices are fed to a probabilistic multimodal algorithm which then predicts the story breaks. The multimodal algorithm takes account of the previous state of the indices and predicts the present state. It is then compared with the actual present indices and story breaks are determined. The segmented stories are then indexed for easy retrieval of the stories.