A proposal of identical video segment detection and its acceleration method for huge-scale broadcast video archive

摘要

In this article, we propose the identical video segment detection which can be used for an automatic detection of important video segments and their relations from huge-scale broadcast video archives. The identical video segment detection finds pairs of video segments created from the same video materials with minor change by post production. The identical video segment detection is defined by using normalized cross correlation. We aim at applying the method to a broadcast video archive server, which can store thousands hours videos and acquire multiple video streams simultaneously. We investigated acceleration methods in order to continuously perform the identical video segment detection to that video archive. We propose the normalized intensity histogram and show that it sufficiently approximates to normalized cross correlation. We also optimize the generation parameters of the histogram by experiments. The total search efficiency was roughly estimated with employing SR-Tree to accelerate the distance comparison between histograms.