UBC Theses and Dissertations
Video database retrieval based on trajectory analysis Gu, Zhe
The increasing volume of video motivates research on video management. Related topics include video representation, organization, query formulation and search. Much work has been done based on shot segmentation of video, however, the temporal nature of video has not been fully investigated. Our goal is to analyze the motion of objects in video, so that video can be retrieved based on objects' motion it records. We emphasize the importance of motion information in video. A two-level abstraction based on the movement of objects is proposed. The first layer describes the movement of objects with trajectory and speed curves, and the second layer abstracts the curves with geometric features. Points with maximum curvature are extracted, then, based on these points, feature values are calculated that are insensitive to planar translation, rotation and scale. Similarity of trajectories is measured by the similarity of the feature values. A similarity measurement metric with point warping is given. We implemented the R-tree as the indexing structure and specialized it to our application. The trajectory and speed curves are analyzed and the abstract feature values are organized within the R-tree structure. A QBE interface is provided, and hierarchical searching is proposed and implemented. Experiments show that video content retrieval based on motion is useful and feasible.
Item Citations and Data