Research On Gunshot Detection In Public Places

In recent years,terrorist attacks have occurred frequently,and the public security has been highly valued by the society and the state.In addition,with the development of Internet of things,the monitoring equipment has become very popular,which also brings great opportunities for the realization and application of the abnormal event detection system.Compared with the video screen,the sound has a better expression of events with the violence semantic such as gunshot,explosion and cry,meanwhile,the processing of audio is faster than the video,and can overcome many disadvantages in video detection,so the abnormal event detection based on audio has attracted more and more attention.The traditional detection of acoustic events mainly follow the features and methods of speech processing.However,the public place is complex,and conventional features is not sufficient for describing the characteristics of the target event,what's worse,the distortion caused by noise also seriously affects the robustness of the detection system.To this end,we starts from the aspects of feature selection and noise reduction,taking the gunshot as the representative event,proposing a set of general algorithm framework,and carried out relevant verification experiments on simulated gunshots in real scenes..In view of the feature representation of gunshots,a feature selection algorithm based on local learning is adopted to solve the problem.And the mixture minimum description length evaluation criterion is adopted to solve the parameter learning problem of the Gaussian mixture model in different feature spaces.Finally,considering the feature dimension and detection performance,selecting a group of the most representative characterist ics of gunshots,not only improves the distinction of shooting and background noise,but also eliminates the time-consuming extraction of irrelevant features.Aiming at the serious performance degradation of the gunshot detection system under complex real scene noise,the noise reduction of the gunshot is processed by using the non-negative matrix factorization.Considering the locality of gunshots and background noise in the spectrogram structure,a mixture local dictionary is used to more accurately chara cterize them.Thus,the better noise reduction is achieved,and the robustness of the gunshot detection system is improved...