Forensic Audio Analysis and Event Recognition for Smart Surveillance Systems

In this study, we propose forensic audio analysis and event recognition framework for smart surveillance systems. Surveillance systems are getting a growing attention. Integration of audio surveillance systems to IP cameras offers advantages over video surveillance systems solely. For instance, an audio surveillance system can provide an additional information about the exact moment of an incident that might be occurring out of the range of surveillance cameras. In our work, mel-frequency cepstral coefficients (MFCCs), energy, pitch range (PR), and linear prediction coefficients (LPC) feature sets were extracted and evaluated on an open-access DASE database, which contains nine audio events including glass breaking, dog barking, scream, gunshot, explosions, police sirens, door slams, footsteps, and house alarm sounds. K-nearest-neighbor (KNN), support vector machines (SVMs), Gaussian Mixture Model (GMM), and classifier fusion classifiers are designed. All classifiers achieved promising accuracy rates.