Scene boundary detection by audiovisual contents analysis

Scene boundary detection is an essential research in content-based video summary, retrieval, and browsing. In this paper, we present an efficient and robust scene extraction algorithm. The proposed algorithm consists of three stages. The first stage is shot boundary detection, and the second stage is the musical scene boundary detection through detection of musical shot. In the last stage, scene detection among non-musical shots is accomplished. In order to detect musical shots, audio categorization is accomplished on audio clips that are divided into visual shot unit. Then low level audio features are calculated for categorization of audio clips. Finally, the parts of video which are containing music component are discriminated on the assumption that the shots in a scene contain same background music. In scene change detection among non-musical shots, distance matrix among shots is calculated based on visual information and time distances between each shot. To provide a reasonable limitation of time distance, variable length time-window method is proposed. The scene boundaries are detected by using shot clustering and scene formation.