A novel metric for efficient video shot boundary detection

With the current rate of video data generation, there is an urgent need of automatic video content analysis for subsequent purposes such as summarization, retrieval and classification. And video shot boundary detection is usually the first step to segment a video clip into meaningful shots. Taking the processing speed into account, most state-of-the-art methods make use of the frame histogram to extract shot boundary characteristics. In this paper we propose a different approach with a novel metric, which essentially captures the observation that within any shot, a pixel value in any frame usually has a pixel value very close to it within a small neighborhood in an adjacent frame. It turns out that the proposed approach can make better use of frame structural content than the histogram approach. In addition, the proposed metric has a low computational complexity. We propose a video shot boundary detection algorithm based on the proposed metric for detecting both cut transition (CT) boundary and gradual transition (GT) boundary. Experimental results show that the proposed approach enjoys better detection rates over the state-of-the-art with competitive processing speed.