Video summarization using feature dissimilarity

This paper presents a novel video summarization method using iterative shrinkage of potential scene feature dissimilarities. The proposed method starts by segmenting the initial video into peak frames having sharp scene variations for summarization length prior. Then, visual significant frames are estimated using visual motion dissimilarities between input video frames and peak frames.

I. Introduction

The representation of videos as a sequence of a large number of consecutive frames presents numerous limitations in surveillance video data for fast content-based search, retrieval, navigation, and storage. It is important to segment the video into homogeneously clustered segments in content space and then to describe each segment by a compressed and sufficient number of frames to obtain a video summarization [1] [2]. For this reason, the most significant frame selection method is required for efficient video summarization.