Designing video data management systems
Arun Hampapur · Deep Blue (University of Michigan) · 1995
Video is the most effective means of communicating and storing audio visual information. The usability of video acquires a new dimension when it is augmented with content based access. Video data management systems manage video collections while providing content based access. The design of such systems is the topic of this thesis. Managing video data adds a new dimension of complexity to the data modeling, insertion, organization and retrieval tasks. The modeling and insertion of video are the focus of this work. Video data models are application specific representations designed to facilitate typical data access patterns. A set of video applications are analyzed to arrive at the design of a data model. The data model uses a temporally segmented representation of video with attached descriptions called features. Multimedia authoring and video production is used as an example application to design video insertion procedures. Video segmentation comprises the first task in video insertion. Segmentation is formulated as feature based classification. The design of the feature detectors is based on video production models derived from standard movie and video production techniques. Feature detectors for effects like fades, dissolves and page translates have been designed. A discriminant function combines the feature detector results to achieve video segmentation. Video indexing comprises the second task in insertion. This is formulated as feature based classification. A methodology for designing feature based indexing schemes are proposed. Indexing schemes are ranked based on an efficacy measure. The indexing scheme uses computational constrains and poses classification as a design problem. An image motion based video classification is used to design a cinematographic video indexing scheme. The segmentation and indexing algorithms are incorporated into a prototype system which operates on video from commercial cable television. The thesis presents results of experiments with half hour of video data.