Content-based Video Structural Analysis

Video structural analysis is the process of converting the video sequences of two dimensions to structural data, which is the base of content-based video browsing and retrieval. It mainly includes two parts: temporal video segmentation and key-frame extraction. In this paper we classify most existed temporal video segmentation methods according to dissimilarity measure and threshold specification. Then we analyze their shortfalls. We measure the discontinuity value using MPEG-7 descriptors and take into account additional information and prior information, which includes variance pattern and shot-length distribution. With above thoughts, this thesis makes progresses in following aspects:(1) Using the Scalable Color Descriptor of MPEG-7 as feature to measure the discontinuities, which has the advantage of histogram-based method that is insensitive to object motion and besides improves the interoperability of various systems.(2) Improving the existing abrupt cuts detecting mechanism, which can filter the effect of extreme factors and increase the accurateness.(3) By borrowing experiences from the existed gradual transitions detection mechanism, we propose a novel algorithm for detecting gradual transitions, which not only identifies the start of gradual transition but also the end of it. Moreover, it also can classify gradual transitions.(4) Extracting the key frames based on motion descriptor of MPEG-7, which works much better than selecting the previous approaches.We develop the prototype of above video structural analysis system and make experiments based on different kinds of video sequences. Experiment shows the following facts: the proposed method for video segmentation is robust and interoperable and the key frames extracted by MPEG-7 motion descriptors are much more representative than the previous approaches.