Research On Key Technologies Of Perimeter Alertin The Camp Area Based On Audio-video

In recent years,emergencies and abnormal events have been increasing year by year.Facing with these problems,it is more and more difficult for traditional artificial monitoring system to meet the security requirements.Especially for some security sensitive areas,such as the perimeter area of the prison and military camp,for safety management requirements,detection of personnel intrusion in surrounding areas is required.When suspicious targets enter the area,the monitoring and control system can automatically analyze and make a decision,and it alerts to notice the security personnel in time when there is an emergence situation.So intelligent video monitoring system becomes the main solution,but because of the complexity of the monitoring scenario,it is likely to result in high false alarm rate and missed detecting rate when the system only relies on the video image information for intelligent monitoring.The abnormal events are often accompanied by abnormal sounds,and in some cases an audio signal even contains more direct information than a video signal.Therefore,we can introduce the audio information to make up for the lack of video monitoring,increasing the control of monitoring scene information.In view of the application background above,two aspects of research,the pedestrian detection in video and the abnormal sound recognition in audio,are carried out in this thesis.The main research work is as follows:Firstly,based on the traditional pedestrian detection framework of RGB channel template and convolution neural network,a pedestrian detection method combining multi-channel template and convolution neural network is proposed.The method uses motion detection algorithm to detect the region of interesting,so that the pedestrian detection efficiency can be improved and the false alarms can by reduced.Meanwhile,a multi-channel template and maximum fusion strategy are adopted to reduce the possible pedestrian missed detection which is caused by the RGB channel template based algorithm.Through the pedestrian intrusion detection experiment,the effectiveness of the above method is verified.Secondly,aiming at the identification of abnormal sounds such as gunfire and explosions,we propose an abnormal sound identification method based on M-MFCC features and paired SVM.On the one hand,we use Hilbert transform instead of Fourier transform to improve the performance of MFCC feature extraction method,enhancing the significance and discrimination of the characteristics of short-time high-energy sound signals.At the same time,the frequency spectrum transformation and the filtering method are independently performed by the front and back half frame sound signal,reducing redundant computation and improving the feature extraction efficiency.On the other hand,we designed a cascading binary SVM classifier to realize the classificationof three kinds of voice signal,resolving the low operational efficiency of multiple targets classification.The comparison test shows that the proposed method has low false rejection rate and false accept rate,which can be used as the effective complement of video monitoring.