The Design And Implementation Of A Video Surveillance Abnormal Behavior Recognition System Based On Deep Learning

In recent years,with the development of video technology and the widespread use of high-definition equipment in more and more fields,behavior recognition technology based on this,especially abnormal behavior recognition,has also a broader development space,In special engineering scenarios,an effective abnormal behavior detection system can greatly save resources and prevent problems before they occur.This topic originates from a monitoring project of abnormal behavior of a substation of an electric power company.Based on the monitoring video,the workers and pedestrians in the substation are tracked and identified for smoking and phone calls.Once found,warning messages are issued and stopped in time to prevent accidents.First,the domestic and foreign status of behavior recognition technology is introduced and compared.Several mainstream target detection algorithms and their advantages and disadvantages are analyzed.The RPN in Faster-R-CNN is very effective in extracting pedestrian candidate regions.Therefore,this method is selected to pre-process video frames.At the same time,due to the specificity of engineering scenes In order to track the pedestrians in the video,a multi-target tracking method based on the Kalman filter equation is selected to be used in this system;in order to determine whether the detected pedestrians have abnormal behaviors such as making phone calls or smoking,it is necessary to perform pedestrian actions Recognition and classification,compared the dual-stream convolution method of HOG + SVM and TSN,decided to use deep learning for feature extraction and classification,and finally chose C3 D network as the basic training for classification training for smoking behavior,and introduced RPN to generate candidate regions.Mechanism,a network model of R-C3 D shared network parameters is constructed to achieve end-to-end detection;ResNet network is determined as a classification training network for calling behavior.Using the method of target detection in the system,About 27,000 datasets with three action classifications were processed,and an abnormal behavior recognition model was obtained through tensorflow training to implement a calling and smoking behavior classification system.Recording and collecting videos of different scenes and complexity to test the system,the results show that the accuracy of pedestrian detection is more than 98%;the accuracy of pedestrian tracking in complex scenes with multiple targets is about 85%;the accuracy of abnormal behavior recognition can finally reach About 84%;In terms of processing speed,it basically meets the real-time requirements and can be used in actual projects.