On efficient clustering and indexing methods over large video sequences

Abstract

As video data is widely used over the Internet, how to manipulate video efficiently is a challenging problem to many researchers. In order to perform content-based video similarity search, we need to firstly perform video summarization to extract the representative frames from each video clips, namely, keyframes. The similarity between two video sequences can thus be measured by the percentages of similar keyframes shared by both video sequences, when each keyframe is extracted and represented as a highdimensional feature vector. The direct computation of such similarity over the whole dataset would, however, involve time-consuming sequential scans over the entire highdimensional data space. Meanwhile, existing indexing techniques on high-dimensional vector space suffer from the notorious “curse of dimensionality”. Therefore, an efficient and effective indexing method is needed to reduce the computation cost for the similarity search on video sequences. In this thesis, we propose an efficient subspace discovering and clustering technique to automatically discover correlated clusters. We adopt a dimensionality reduction technique called Principal Component Analysis (PCA) to transform the vector data into a new vector space with reduced dimensions, and iteratively discover clusters on the deeper subspaces. A corresponding indexing structure is proposed to reflect the cluster hierarchy, so that we can then easily prune out large potions of dissimilar frames by looking at the relationship of their corresponding clusters. Extensive experiments over large-scale video datasets show the efficiency and usefulness of our method.