Video Shot Detection Using Hidden Markov Models with Complementary Features

Abstract

Shot detection is the first stage of video analysis. In this paper, we present a machine learning based shot detection approach using hidden Markov models (HMMs), in which both the color and shape clues are utilized. Its advantages are twofold. First, the temporal characteristics of different shot transitions are exploited and an HMM is constructed for each type of shot transitions, including cut and gradual transitions. As trained HMMs are used to recognize the shot transition patterns automatically, it does not suffer from any trouble of threshold selection problem. Second, two complementary features, statistical corner change ratio (SCCR) and HSV color histogram difference, are used. The former summarizes the shape well whereas the latter summarizes the appearance well. Experimental results on a set of test videos demonstrate the efficacy of this shot detection approach