Automated annotation of significant scenes based on structure of broadcast sports video

For broadcast sports video, the authors attempted automatic generation of semantic annotations for parts that seem important from the viewpoint of story, in accordance with the composition of the video and the game. First, the parts of a sports video which seem significant in the progress of the game are extracted from the closed caption text stream by keyword sequence search. The information concerning the play and the player in each part is extracted and the, annotation is generated. Then, in order to determine the position in the video to which the annotation should be attached, the video is partitioned by similarly extracting the game-progress part by matching to the image stream. Finally, by establishing the time synchronization of the two streams, the generated annotation is attached to the video. An experiment was performed in which the method was applied to videos of American football games, as an example of sports videos. Automatic annotation was achieved accurately for the game-progress part with a reproduction rate of 75% and a match rate of 90%. (C) 2003 Wiley Periodicals, Inc.