Content-based video retrieval (CBVR) is now a researchhotspotin multimedia discipline, it integrates the content of digital image processing, digital video processing, multimedia technique, database technique etc. and uses the video features as a retrieval condition to overcome the shortcomings of traditional retrieval methods based on attributes. In this dissertation, a deeply research is carried out on four key issues: video shot segmentation, feature extraction, feature describtion and modeling, retrieval algorithm.In video shot segmentation, an improvement to double-threold shot segmentation algorithm is provided, which uses multi-frame sampling technique and can improve the performance significantly on the detection of gradual transition. An abrupt transition detection algorithm is also developed on the basis of the closest pixels matching in spatio-temporal slice, which decreases the false rate and computing strength greatly.In feature extraction, a difference motion detection algorithm of adaptive background updating is proposed, which has a good robustness to the changing of background and illumination.In feature description, a novel topological relationship set which is composed of six named topological relationships and an inverse operator is defined, and its completeness and exclusiveness are proved. Then, a video shot model based on MPEG-7 standard is constructed and a description scheme to video shot is implemented perfectly.In retrieval algorithm, a row-column coded retrieval algorithm is developed, which is robust to the frame shift and illumination changing. A novel hierarchical architecture to organize different algorithms in content-based retrieval is proposed. Experiment shows that this architecture improves the retrieval efficiency.According to the achievements above, an experiment system is designed and implemented. It can carry out conventional content-based video retrieval, and can also search special video clip in real-time video stream.