In this paper, we investigate the similarity search methods for large video data sets that are the collection of video clips. A video clip, a sequence of video frames describing a particular event, is represented by a sequence in a multidimensional data space. Each video clip is partitioned into video segments considering temporal relationship among frames, and then similar segments of the clip are grouped into video clusters. Based on these video segments and clusters, we define similarity functions and present two similarity search methods: the HR (hyper-rectangle)-search and the RP (representative point)-search. Experiments on synthetic sequences as well as real video clips show the effectiveness of our proposed methods.
展开▼
