Baseball video indexing using patternization of scenes and hidden Markov model

This paper proposes a method to automatically set indexes which describe the content of a baseball video scene. The approach is based on the technique to patternize baseball scenes using a set of rectangles with some image features and a motion vector. Each baseball scene is expressed as a sequence of symbols based on the patternized data of every shot and given "event-indexes" (home-run, single-hit, walk, etc.) using hidden Markov models which have been trained to the sequences. The processes to construct training data and to set event-indexes to baseball scenes are described in detail. The proposed method is evaluated by an experiment using seven major league baseball video games.