One-class SVM based approach for detecting anomalous audio events

The last generation automated security and surveillance systems call for new and advanced capabilities to automatically and reliably recognize suspicious events or activities in the monitored environments on the base of a real-time and combined analysis of different multimedia streams. In this paper we focus our attention on the analysis of audio signal and present a method based on one-class Support Vector Machine (1-SVM) classifiers. Such an approach is able to support the recognition of different kinds of burst-like anomalies (i.e. gun-shots, broken glasses and screams), on the base of their time and frequency domain characterization. Several experiments have been carried out, showing the potentiality of our method with respect to other approaches proposed in the recent literature.