A new content-based scene change detection method on compressed video

Scene change detection is the first step for automatic indexing of video data. In particular, it is desirable to detect scene changes on compressed data because this requires less processing overhead and storage to hold the results. Some approaches are proposed to detect scene changes on compressed data by computing the variances of DCT coefficients. However, these approaches do not consider the content in the image data, so undesirable results are often generated and no semantic information about the content is available. We propose a content-based scene change detection method by finding 3D connected volume from compressed video data. First, we find meaningful regions on the reduced DC image sequence using spatial coherence. Then, we compute correspondences of meaningful regions between successive frames using temporal coherence and finally detect scene changes by computing the change rate among meaningful regions along the time direction.