HAH-tree: towards a multidimensional index structure supporting different video modelling approaches in a video database management system

This paper proposes a multidimensional distance-based index structure for video data which supports the three important video modelling approaches namely hierarchical unit-based modelling, feature-based modelling and video semantics modelling seamlessly within one single framework. These three modelling techniques collectively capture and contain the important aspects of the users' information need during content-based video retrieval. The index is built based on the low-level features of the video data, and the hierarchical containment relationships as well as the video semantics are introduced into the index space with an efficient data signature and a stochastic model, respectively. Efficient k-NN algorithms are proposed to emulate popular content-based video retrieval approaches in a multidimensional distance-based index structure. Extensive experimental results demonstrate the capability of the index structure to generate relevant query results with low computational overhead.