中文摘要：

为方便视频的浏览和存储,以概括视频内容为目的的视频摘要技术变得十分重要.针对目前在视频摘要问题中,根据先验知识事先确定和阈值调整2种关键帧数目的确定方法的灵活性及直观性不足这一问题,提出一种基于SVD和稀疏子空间聚类的视频摘要算法.该算法利用SVD对视频在时间维度上进行维数缩减,以累计贡献率为度量标准确定视频在时间维度上的主成分分量个数,将主成分分量个数作为关键帧数目;然后利用稀疏子空间聚类算法对视频帧进行聚类;最后在每一类中选取与其他视频帧相关性最大的帧作为关键帧,生成视频摘要.实验结果表明,文中算法生成的视频摘要内容覆盖率高,可以根据视频长度和类型灵活直观的确定关键帧数目,并且对于不同类型和长度的视频给出了累计贡献率的取值范围,可以为用户提取合适长度的视频摘要提供有效依据.

英文摘要：

In order to make video browsing and storing more convenient, the technique of video summarization whose purpose is summarizing the content of video becomes very important. At present, keyframe numbers determination method can be organized into two categories: the one determining the numbers according to the priori knowledge, and another according to threshold adjustment. However, flexibility and intuition are mostly lacked in both methods. To solve this problem, a video summarization method based onsingular value decomposition and sparse subspace clustering is proposed. Firstly, time dimension of the video is reduced by singular value decomposition, and principal component numbers on time dimensionwhich are determined by the cumulative contribution rate are regarded as the keyframe numbers. Secondly,the video frames are clustered by sparse subspace clustering. Finally, in each cluster the frame which has thebiggest correlation with other frames is selected as keyframe. Experimental results indicate that the proposed method can generate video summarizations with high content coverage rate, and adjust the lengths of video summarizations flexibly and intuitively according to lengths and types of videos. The range of the cumulative contribution rate of videos with different types and lengths is given, which can provide an effective basis for users to extracting a summary with an appropriate length.