Verification of the Parameterization Methods in the Context of Automatic Recognition of Sounds Related to Danger

Digital signal processing of sound is a domain with numerous applications in the telecommunications and informatics. These well-developed algorithms of the analysis of sound can be also applied in the field of security systems, where traditional monitoring is still based mainly on video cameras. The commonly used monitoring cameras can be equipped with additional microphones and the audio content can be analyzed by a monitoring program running on a dedicated hardware. This application can automatically detect in the audio stream events like a broken window, gunshot, explosion, or scream. One of the main parts of this system is a parameterization block. In this article two parameterization methods are proposed for this purpose. The first is based on the frequency analysis of the examples of the sound events. The second is based on using a standardized set of audio MPEG-7 and cepstral descriptors. The feature vectors calculated by these two methods have been used for the training of two intelligent classifiers: a support vector machines classifier (SVM) and a neural networks perceptron (NNP). The classifiers have been verified using of the cross-validation method. The results have been compared and conclusions derived. The application of the results in a system working in real conditions is presented and discussed at the end of the article. The work has been done in the frame of the international project “INDECT” (Intelligent Information System Supporting Observation, Searching and Detection for Security of Citizens in Urban Environment).