Classification of Video Sequences in MPEG Domain

Abstract

This chapter describes a method for automatic classification of video shots from a video database by using distance metrics derived from motion information only. The classification serves as the first step of the indexing process of a video scene and its retrieval from a large database in order to partition the database into more manageable sub-units according to the types of scenes, e.g. sport, drama, scenery, news reading. The method is intended for web-based and telecommunication applications and therefore the processing is carried out in the MPEG (compressed) domain making use of the spatio-temporal data already available in MPEG video files. The confidence of the MPEG motion vectors estimated by the block matching algorithm is evaluated using a block activity factor, for retaining or discarding the vectors from the classification distance measure by a filtering process of the MPEG motion vector fields. The chapter presents a robust regression technique, based on Least Median-of-Squares, to deal with the situation. A novel metrics called activity power flow is introduced to effectively capture the spatiotemporal evolution of scenes through the video sequence.