Abstract

A system for recognition of threatening acoustic events employing parallel processing on a supercomputing cluster is featured. The methods for detection, parameterization and classication of acoustic events are introduced. The recognition engine is based onthreshold-based detection with adaptive threshold and Support Vector Machine classifcation. Spectral, temporal and mel-frequency descriptors are used as signal features. The algorithms are implemented in a supercomputing environment utilizing a specializedframework for processing multimedia data streams. The recognition engine is evaluated in various conditions, both using pre-recorded signals and real-world events. First, an evaluation in laboratory conditions is performed to simulate selected acoustic environ-ments and evaluate the recognition rates in noise. Subsequently, the recognition rates are assessed in various practical situations (related to indoor and outdoor surveillance)and compared with the results obtained in simulations. The adaptation of event detection evaluated by comparing dierent approaches to adapting the detection thresholds. Finally, parallel processing is introduced to improve the performance of the developedrecognition engine. The experiments utilizing a supercomputing platform are introduced, which show that the employment of parallel processing leads to signicant shortening of the time required to make the decision. The possible practical applications ofthe developed methods are outlined, including surveillance of urban space, public events or private property.

Author (1)