Tools for Advanced Video Metadata Modeling

Abstract

In this Thesis, we focus on problems in surveillance video analysis and propose advanced metadata modeling techniques to address them. First, we explore the problem of constructing a snapshot summary of people in a video sequence. We propose an algorithm based on the eigen-analysis of faces and present an evaluation of the method. Second, we present an algorithm to learn occlusion points in a scene using long observations of moving objects, provide an implementation and evaluate its performance. Third, to address the problem of availability and storage of surveillance videos, we propose a novel methodology to simulate video metadata. The technique is completely automated and can generate metadata for any scenario with minimal user interaction. Finally, a threat detection model using activity analysis and trajectory data of moving objects is proposed and implemented. The collection of tools presented in this Thesis provides a basis for higher level video analysis algorithms.