Video partitioning and video segmentation in content-based video databases

A novel multi-resolution video partitioning algorithm, called MVP, is introduced. MVP improves the efficiency of video partitioning by using a twin-step search process. A larger step is used until a shot boundary is suspected. Then, a finer step is used to find the type and exact location of the shot boundary. The method is applicable to both uncompressed and MPEG compressed video streams. It may be used to detect both abrupt and gradual shot boundaries. In addition, an automatic localized threshold selection algorithm is provided. The threshold is adaptively set based on the local characteristic of the filtered difference signal.The unsupervised video segmentation technique proposed in this thesis works at object level, and aims to extract moving object sequence from the video. The proposed method works in two stages: motion segmentation and moving object tracking. In the first stage, the frame is segmented into temporally homogeneous regions. A robust motion segmentation algorithm based on Multiple Structure Robust Estimator (MSRE) is presented to achieve the task. In the second stage, we first identify moving objects by merging the temporally homogeneous regions on the basis of spatial similarity measure. The models for the objects are then constructed automatically. A model based tracking algorithm is used to track these models in the subsequent frames