Acoustic signal based abnormal event detection in indoor environment using multiclass adaboost

This paper addresses the problem of abnormal acoustic event detection in indoor surveillance systems related to safety and security. The proposed concept event detector determines if the acoustic state is either normal or abnormal from accumulated series of acoustic signals using MFCC and deltas coefficients as acoustic feature vectors and a multiclass Adaboost based acoustic context classifier. A novel concept of adopting an exponential criterion and weighted least square solution to boost binary weak classifiers is proposed here for performance and speed improvements over the conventional and prominent GMM based classifiers.