Developing an integrated video analysis system

Abstract

Matching the similarity between two units of data occurs as a frequent task in video or image analysis. The parameters of matching techniques are level of abstraction of features, distance measures and normalization of features, if supported, or else the method of relatively weighing the features. Most multimedia analysis systems employ only low level features with distance measures similar to Euclidean distance, with no method to automatically generate the weights of the features and thus are ineffective in replenishing suitable matches to the user's demands. We argue for shifting the burden of mapping the feature space with relevant categories from the user to the multimedia analysis system. In this paper, a Bayesian Framework is presented where the evaluation of the parameters of classification and especially the relevancy of each feature with respect to each class is performed automatically. The probabilistic framework is extended to work well for generalized multi-modal distribution of a particular class over the feature space. Theoretical foundation is developed to provide simultaneously existing multiple views to an image or a video sequence. The low-level features can be synthesized with intelligent association to furnish high-level features, which could be more meaningful to the user. The significance of this work is presented by comparing with a system which employs a unsophisticated approach similar to common systems where feature vector of query image and feature vector of template image are compared by means of weighted Euclidean distance. The superiority of our approach is presented over the database consisting of 300 video sequences comprising of diverse video classes.