A hybrid CSVM-HMM model for acoustic signal classification using a tetrahedral sensor array

In this paper, we propose a new framework for classification of multi-channel acoustic signals collected with a tetrahedral sensor array from the events of launch and impact of two different weapon types. While the temporal dynamics of transient acoustic signals can be well represented by Hidden Markov Models (HMMs), HMM is not an effective discriminative model. On the other hand, discriminative models such as Support Vector Machines (SVMs) are not able to capture the useful dynamic information and cannot handle variable length features obtained from temporal signals. Thus, in this work, we integrate SVMs and an HMM-based representation model to improve the classification of the four events. Moreover, Contextual SVMs (CSVMs) are employed in this system in order to capture the higher-order correlations among the multiple channels of the acoustic signal, and also handle the situations, where one or more of the four sensors fail. Experimental results indicate that the proposed model results in significant improvement in classification accuracy of the multi-channel acoustic signals compared to a traditional HMM framework and SVM classifier.