Temporal Bayesian Network based contextual framework for structured information mining

Specific domains in video data contain rich temporal structures that help in classification process. In this paper, we exploit the temporal structure to characterize video sequence data into different classes. We propose the following perceptual features: Time-to-Collision, shot length and transition, and temporal motion activity. Using these perceptual features, several video classes are characterized leading to formation of high-level sequence classification. Resulting high-level queries are more easily mapped onto the perceptual features enabling better accessibility of content-based retrieval systems. Temporal fusion of the perceptual features forms higher-level structures, which can be effectively tackled using the Dynamic Bayesian Networks. The Networks allow the power of statistical inference and learning to be combined with the temporal and contextual knowledge of the problem. The modeling and experimental results are presented for a number of key applications, like sequence identification, extracting highlights for sports, and parsing a news program.