Audio Content-based Highlight Detection Using Adaptive Hidden Markov Model

A hidden Markov model (HMM) is employed in this paper to model the structure, and thus generate highlights, of table tennis games. Unlike existing approaches, we use only one HMM for the table tennis games, which combines domain knowledge during initialization, and can be easily adapted to new games through a training phase. Both the homogeneity and heterogeneity are taken into account during this modeling. With this HMM embedded, a multi-level audio content-based highlight detection system, incorporating four audio keywords that are created according to time and frequency domain low-level audio features, is proposed to perform highlight detection in table tennis games. Encouraging results are obtained in the experiments carried out on six table tennis matches