Regional Management Intelligent Video Analysis System Based On Deep Learning

With the development of technology and the progress of society,video surveillance system has been an important security guarantee in people's daily life.However,with the recent explosive growth in the number of cameras,traditional video surveillance systems only record but the unprocessed features are increasingly unacceptable.While too much manpower and material resources are spent on it,there are still inefficiencies,false alarms,missed alarms,staff fatigue and other shortcomings.Therefore,how to intelligently,automatically and efficiently monitor the monitoring system has always been a hot research issue in the engineering field.In recent years,with the rapid development of artificial intelligence technology and computer science,although deep learning technology is the core of technology,breakthrough progress has been made in the field of computer vision.This article compares various deep learning algorithms,and uses the more advanced Yolov3 algorithm in deep learning to designed an intelligent people regional control system with pedestrian detection as the core,supplemented by subsequent video analysis processing and database storage functions.The system can avoid the dependence on labor in the traditional monitoring system,improve work efficiency,and ensure the stability and accuracy of the intelligent analysis function.This paper first builds the software and hardware architecture of the intelligent people regional control system based on deep learning.The system hardware consists of two parts:the monitoring system and the access control system.The two parts are connected to the same processing host through the full Gigabit switch.The system software obtains the video information transmitted from the hardware system,and uses the fine-tuned YOLOv3 to analyze and detect the video stream and combined with the database to store the processing results.The system design is reasonable,the work is efficient,and the operation is stable in the field test.Due to the scenes applied in the system,there are significant differences between the personnel target and the personnel targets in the public data set in terms of shooting angle,distance,and occlusion.This paper also constructs a sample set suitable for the detection system of this paper by using the video image information collected in the field,and fine-tunes the original YOLOv3 model under the guidance of migration learning,and then the fine-tuned YOLOv3 model was compared with original YOLOv3 model.The experimental results show that the fine-tuned YOLOv3 model has advantages over the original model inthis experimental scenario.The detection speed reaches 29ms/frame,and the false alarm rate and the false alarm rate are reduced to 1/5 and 1/2 compared with the original YOLOv3 model before fine-tuning,reaching 0.3% and 3.6%,which is obtained while ensuring real-time and detection performance.