The Study Of Abnormal Sound Recognition For Security Monitoring

The monitoring surveillance, which is widely applied, is an important technology method in the field of public security early warning. The abnormal sound can effectively reflect abnormal situation in the scene from the perspective of auditory monitoring acoustically, making up the disadvantage of monitoring surveillance. Abnormal acoustic recognition in public place can effectively warn the occurrence of major accidents and emergency, thus it is followed with wide interest by domestic and overseas research institutions. Based on studying lots of articles and papers, the following researches are carried out about abnormal acoustic recognition in public place.Abnormal acoustic database in public place was built up in consideration of a deficiency of abnormal acoustic database source. Firstly, abnormal sound is introduced. Then, according to the occurrence frequency and harm level of abnormal sound in public place, the data of abnormal sounds produced by gunshot, exploiting and breaking of glass etc. are collected to make up with the abnormal acoustic database record. Finally, the time-frequency features of these abnormal acoustic data are analyzed.In the light of the low abnormal acoustic classification rate problem caused by the security surveillance scene complexity and abnormal acoustic signal non- stabilization, this paper put forward an improved MFCC which is based on MVDR spectrum estimation combined with short-term energy for abnormal acoustic feature extraction algorithm. The proposed algorithm used MVDR spectrum estimation to replace FFT spectrum estimation which is contained in traditional MFCC to extract MVDR-MFCC feature, then expressed abnormal acoustic from different direction combined with short-term energy. Simulation experiments show that the new algorithm has better recognition performance, the average recognition rate is increased by 2.5%, compared with traditional abnormal acoustic feature extraction method, it can effectively classify acoustic breaking of glass, screaming, explosion, shootings, etc and is more suitable for real application scenario.In accordance with problem of low robust of recognizing abnormal acoustic signal based on MFCC feature in the complicated security surveillance scene, this paper put forward an abnormal acoustic feature extraction algorithm which is based on Hilbert spectrum estimation and independent component analysis(ICA). Referring to MFCC, the power spectrum of Hilbert spectrum estimation result is calculated firstly and then logarithm of the power spectrum filtered by Mel filter bank is taken, finally de-correlation, dimensionality reduction and noise reduction by ICA is achieved. Simulation experiments show that the new algorithm has better system performance, and effectively raised robust of abnormal acoustic classification system compared with MFCC method.