3 classes segmentation for analysis of football audio sequences

Abstract

We are dealing with segmentation of audio data in order to analyse football audio/video sequences. Audio data is divided into short sequences (typically with duration of one or half a second) which is classified into several classes (speaker, crowd and referee whistle). Every sequence can then be further analysed depending on the class it belongs to. In order to segment audio data, several methods are presented. First simple techniques are reviewed for segmentation in two classes. From the limitations of these approaches, a method based on cepstral analysis is detailed. Next we present two more complex methods dealing with 3 classes segmentation. The first one is based on hidden Markov models whereas the second one is a combination of a C-mean classifier and multidimensional hidden Markov models.