This thesis examines the problem of video segmentation and summarization from a results fusion perspective. Many techniques have been developed for the segmentation and summarization of digital video. The variety of methods is partially due to the fact that different methods work better on different classes of content. Global histogram-based segmentation works best on color video with clean cuts and global intensity changes; local histogram-based segmentation is less sensitive to region changes in the video and therefore works better when scenes consisting of similar content are shot from different angles; DCT-based segmentation algorithms attempt are less sensitive to abrupt intensity changes due to lighting effects such as camera flashes; edge-based segmentation algorithms work well when high quality edge information can be extracted from the video sequence, motion-based summarization works best on video with moving cameras and a minimum of disjoint motion. Results fusion combines the properties of these varying algorithms into a common framework that can benefit from the advantages of each disparate approach. Recognizing that there is no single best solution for each of these problems has led to this work in integrating the variety of existing algorithms using results fusion methods.; The work is divided into four parts. The thesis begins with an in-depth study of the various video segmentation methods. This chapter categorizes the existing shot segmentation and summarization methods, noting their strengths and weaknesses. Next, results fusion based algorithms and implementations from a variety of fields are reviewed and studied so as to understand the methods that can be applied to video segmentation and summarization. This chapter examines results fusion research from the document retrieval and biometric communities and with an eye towards application to the video domain. The third part of this work presents the results of applying results fusion for video segmentation. This section compares and contrasts individual algorithms with the results fusion implementations. Finally, it is demonstrated that the results fusion methodology used for video segmentation can be extended to video summarization.
展开▼