A video retrieval system based on compressed data from MPEG files

The paper describes a prototype video database system which makes use of MPEG video files to extract content parameters such as colour, shape, and motion vector statistics of regions or objects in video for indexing and retrieval. The content-based video retrieval system uses a visual-sketch approach for the design of the graphical query interface. The system uses a two-stage indexing and retrieval approach - first to classify video shots into separate databases using their motion characteristics, and second to search each database for colour features. The two-stage approach reduces the search time to almost half at a slight reduction in retrieval rate. It is found that colour (together with shape) is the most effective attribute, followed by motion, for video indexing and retrieval applications.