Digital video analysis and manipulation

More and more digital video is being generated and becoming ubiquitous in many emerging applications. Due to the inherently large amount of data, managing and accessing the video according to content in large video databases is very difficult and time-consuming. Efficient ways to analyze, manipulate, annotate, browse and retrieve video data of interest are therefore important and have attracted much attention.; In this thesis, new techniques are proposed to analyze and manipulate the content of digital video. As motion is a fundamental constituent of the visual information in video and often serves as an important clue for understanding the video content, algorithms are developed to compute the dominant motion from images. The computation of this image motion requires the establishment of some form of correspondences between the images. By formulating the measurement of feature correspondences as a parameter estimation problem, robust algorithms are developed to select good image features for establishing image correspondence. The proposed image motion estimation techniques are then applied to several useful applications such as constructing large field of view panoramas from image sequences, stabilizing and smoothing video sequences with unsteady content, and detecting and segmenting moving objects of interest from video sequences.; Next, for the purpose of video matching and searching by low-level visual content, proximity indices and metrics, as well as dynamic time warping techniques are proposed to measure the similarity of video content, both spatially and temporally. A framework for searching a video sequence for segments which are similar to an example video segment of interest is presented.; Finally, new techniques are developed to infer high-level content of structured video in compressed domain. By using MPEG data without full frame decompression and incorporating prior knowledge of basketball video, algorithms are designed to identify interesting basketball events such as wide-angle views and close-up views of the game, fast breaks, shots on basket, etc. The techniques developed are useful in annotation, browsing and classification of the video data of interest by high-level video content.