Algorithm Design And Research On Abnormal Sound Recognition In Public Places

Audio surveillance technology is one of powerful tools for public security surveillance,and the key to make it intelligent is automatically detecting and recognizing abnormal events which are often accompanied with abnormal sounds such as explosion,gunfire and scream etc.overlapped in background sounds in most public occasions.Compared with background sounds mostly continuous existing,abnormal sounds tend to be sporadic and isolate with high tone and great intensity.Nowadays,most traditional abnormal sound recognition methods are often restricted to using features of speech signal processing methods whose generating mechanism is usually different from abnormal sounds,which leads to bad description of abnormal sound characteristics.What’s more,they often set up classification models based on abnormal sound samples some of which are so difficult to obtain sufficiently that the classification boundary of the training model is inaccurate and recognition results are bad.Therefore,we do some research on above problems of the abnormal sound detection and recognition in public places,and our main work is summarized as follows:(1)Put forward a non-overlapping statistical equal MEL feature which is adopted by improving Mel Frequency Cepstrum Coefficient(MFCC)with a non-overlapping statistical equal rectangular filter bank.First of all,in order to obtain the maximum information entropy of the feature,a non-overlapping statistical equal rectangular filter bank is built combined with the statistical amplitude spectrum of abnormal sound in MEL frequency domain.And then,we use the filter bank to filter the input sound signal to obtain the energy information of different frequency components as the feature of the input signal.In addition,to further enhance the identification ability of the feature,the energy differential characteristics of various frequency components is calculated and integrated with the energy feature by the normalized multiple features weighting fusion method strategy to form the final non-overlapping statistical equal MEL feature vector.(2)On the basis of the non-overlapping statistical equal MEL feature and Support Vector Machine(SVM)classifier,an abnormal sound detection algorithm library is established.Extract effective sound events with double threshold detection method firstly,and next obtain the non-overlapping statistical equal MEL feature of each sound event,eventually use the SVM classifier for abnormal sound classification.In order to verify the performance of the algorithms library,series experiments are conducted by using four kinds of abnormal sound,four scenarios sound and their synthetic sound clips from the audio monitoring database established by the IITLAB laboratory.(3)Build the intensity level transition state machine(ILTSM)firstly,and then put forward a kind of an intensity level transition probability(ILTP)feature,and finally set up an abnormal sound detection algorithm framework based on the ILTP feature and GMM background model.In the first place,the intensity transition state machine(ILTSM)is set up based on the statistical characteristics of signal intensity and variance of abnormal sounds and scene sounds.And the intensity level transition probability feature vector is generated according to the transition probabilities between different states.Then we set up background models corresponding to different scenes in public places using the Gaussian mixture model(GMM)model.In addition,we determine the type of a frame signal through the likelihood matching between the ILTP feature of the frame signal and the background model directly but not need to do the sound event extraction preprocessing.At last,confirm its ultimate type and detect abnormal sounds by using the information of more continuous frames.To verify the effectiveness of the proposed algorithms,we do some experiments on the IITLAB laboratory audio surveillance database.The experimental results show that,the algorithm can effectively detect abnormal sound in different SNR environments,and its performance degrades as the reduction of SN R.The FAR and FRR of these abnormal sounds except for gunfire in the real environment are all below 5%,and the detection performance of all abnormal sounds are much better than the other relevant algorithms.