Research On Detection Method Of Violent Behaviors In Video

In the field of public security,video surveillance has become an important part of the security prevention system,which can provide managers with real-time,all-round video streams,but it is unrealistic to rely on human resources to supervise the occurrence of violent incidents 24 hours a day,which is both energy-consuming and prone to blind spots,making it difficult to complete the task of real-time supervision.Applying deep learning technology to the field of video surveillance to automatically detect and analyze video streams captured from cameras,identify the presence of violence,and provide real-time alerts,which helps liberate manpower and improve the efficiency of emergency incident handling.Existing violence detection technologies have many challenges in engineering applications,such as complex backgrounds,human occlusion,and insufficient resolution.In this thesis,we propose to use deep learning technology to accomplish the task of violence detection,and study different solutions for different scenes and surveillance videos under different fields of view to achieve accurate and efficient behavior detection.The primary objective of this thesis is to examine the technology used for detecting violence,along with its associated problems.The paper aims to explore the following areas of work.(1)For indoor and other close monitoring scenes,where the human target in the video is large and easy to do fine recognition,this thesis uses a violence detection algorithm that combines human pose estimation detection and motion analysis to cope with this situation.First,the human pose detection model performs joint point coordinate localization on video frame sequences frame by frame to obtain the corresponding skeleton frame sequences;then,the spatial dependency relationship is modeled for different joint points of the same frame,and the temporal relationship is modeled for the same joint point of different frames to form a behavior analysis model with joint spatio-temporal features,so as to achieve the detection of violent behaviors.The experiments confirm that the algorithm shows good detection effect in the close surveillance environment.(2)For outdoor large-field complex surveillance scenes,where human targets are usually small and not easy to locate human key points,this thesis proposes a violence detection algorithm based on target detection and 3D convolution.First,the human targets and their action sequences in the surveillance area are obtained through target detection and temporal feature extraction techniques;on this basis,temporal features are fused for analysis to detect the presence of violent behaviors in the video clips.Experiments confirm that the algorithm achieves good detection results in long-distance surveillance scenes.(3)Based on the Raspberry Pi 4 embedded platform,an intelligent detection and alarm system for violence is designed and developed.The system is divided into two parts: an embedded detection subsystem and a management service subsystem.The former runs on the Raspberry Pi platform and deploys the violence detection model developed in this thesis,which can connect to the surveillance camera,sample the frame sequence from the video stream,and call the model to detect whether there is violence in the video clip;if violence is detected,the detection information is immediately reported to the management service platform.The latter runs in the cloud and is able to receive information on violence reported by multiple detection devices,manage this information,issue alarms in a timely manner,and provide alarm information query services for the management client.Through the deployment and operation of this system,the effectiveness of the algorithms in this thesis and the feasibility of the scheme are verified.