| Summary: | 碩士 === 義守大學 === 資訊工程學系碩士班 === 94 === In this paper, we aim to use motion features to develop a baseball event detection system. In broadcasted baseball videos, there often exist a large amount of repeated scenes (slow motion replay, inning change and pitcher practicing) and redundant scenes (commercial and spectators scenes), which are mainly due to considerations of business and broadcasting. The major goal of this thesis is to eliminate the useless information and to provide highlight video chips which users are interested in.
In this thesis, we first employ local motion vectors between subsequent frames to compute global motion parameters. Next we define motion features using the global and local motion information. The motion features are then used to develop an event detection system which is able to generate high level semantic. The experimental results indicate that accurate event detection can be achieved by using only the motion features, even if there exist some incorrect estimates of motion vectors.
|
|---|
