Ontology-based shot indexing for video surveillance system

Abstract

Indexing video data is very complex due to the enormous size of video files and their semantically rich contents and unstructured format. Therefore, the first step of video indexing begins with segmenting each video into a number of basic processing units. In general, the most widely used basic unit is a shot that is defined as collections of frames recorded from a single camera operation. Basically, there are two different approaches to index video using these shots. One is using low level features such as color, motion, or object. On the other hand, high-level features such as human interpretations are utilized for the indexing. Recently, a hybrid indexing combining these two approaches has been proposed to take advantages of both of them, and shows that dealing with both features has certain advantages. However, it is necessary that further improvements are required for which features can be extracted, and how they are identified. More importantly, further investigation is needed for the seamless integration of both levels for indexing purpose. To address these issues, we introduce ontologybased indexing for video surveillance system. Generally, video surveillance has non-predefined contents unlike movie or drama, and its main purposes are to detect and record abnormal events. Since ontology defines machine-interpretable definitions of basic concept in a domain, we can prevent semantic inconsistencies from human interpretations and apply it to handle surveillance data. Then, we can classify and index video data more precisely by using well-organized ontology. To illustrate the effectiveness of ontology based video indexing, we implement a Video Ontology System (VOS) to classify and index shots in a simple domain. Our experimental results show that the proposed approach is promising. © Springer Science+Business Media B.V. 2008.