Automated analysis of mpeg-compressed video

To keep pace with the increased popularity of digital video as a communication medium, the development of techniques for fast and efficient analysis of video streams is essential. Automatic techniques are also required to manage the resulting large archives of video data efficiently. This dissertation explores the issues and problems in analyzing large amounts of video without human intervention. With an emphasis on speed, techniques have been developed for direct application to MPEG-compressed video. Temporal segmentation and motion analysis of video are described first. The temporal segmentation algorithm runs at speeds of up to 180 frames per second on SIF-format (352 × 240) video, thus, easily enabling real-time analysis. Techniques for performing more complex tasks such as video indexing and retrieval, video representation and visualization are presented in detail. Key frames are first identified from the video clips. Large numbers of features are extracted from the key frames, and are indexed into a low-dimensional space by dimensionality reduction. Extending the dimensionality reduction technique to features from all the frames of the video clip, we obtain a novel approach to video representation and visualization. Finally, we describe techniques for classifying videos into various categories, with emphasis on identifying sports videos in a heterogeneous collection of videos. We also present a simple technique for extracting news story segments from news videos. We further examine sports videos by providing a taxonomy of sports which can be used as a starting point for classification of sports into types such as indoor/outdoor, team/individual, etc., or for deeper analysis of specific sports. As a case study, we present an approach to analyzing ice hockey clips.