Robust Classification of Abnormal Audio using Background-Foreground Separation

Surveillance systems are largely based on visual mode, but acoustic mode of surveillance have gained importance in the recent times. Detection of abnormal sounds can greatly contribute for surveillance purposes. Audio signal processing has always been challenging due to dynamically changing background noise. Audio event recognition rates are to be increased and false positive rates are to be reduced to build more efficient systems. The proposed methodology focuses on reducing false alarm rates by implementing a Background-Foreground Separation module. Separation of foreground events from the background audio is achieved by introducing an Impulsive Sound Detection module. In the classification stage, k-Nearest Neighbor classifier with Mel Frequency Cepstral Coefficients (MFCC) features are employed. The method is validated on a large publicly available dataset which is developed for acoustic event classification. A recognition rate of 94.07% and a low false alarm rate of 1.23% is achieved which is significant improvement compared to the state-of-art approaches.