Intelligent Video Monitoring System Based On Multi-target Pedestrian Tracking

With the development of computer vision and the improvement of people's requirements for video monitoring system,intelligent video monitoring system based on computer vision has gradually become the development direction of intelligent video monitoring in the future.With the help of target detection and multi-target tracking technology,people can study and judge the behavior of pedestrian targets in surveillance video,and give intelligent warning of pedestrian capture and abnormal behavior.At present,the effect of pedestrian capture and early warning in simple scene is better,but for the scene with large pedestrian flow,such as the peak subway station up and down passenger flow,crowded pedestrian street and the entrance and exit during the exhibition period,because of the dense pedestrian,the real-time and performance requirements of the algorithm are higher,which makes the guarantee of real-time and accuracy of pedestrian capture and early warning become an urgent problem to be solved.In view of the above problems,this paper uses detection based tracking(DBT)to research and apply pedestrian detection and multi-target tracking under the fixed camera in video monitoring,and realizes a real-time and high-precision intelligent video monitoring system.The main research contents of this paper are as follows:(1)Pedestrian detection.This paper studies the method and principle of yolov3 detection framework,and completes the production of pedestrian detection data set including 13710 sample images for the road monitoring scene in the project area.experimental verification shows that the m AP of the target detection algorithm under the YOLOv3 detection framework is 0.829.The detection speed of 2-megapixel surveillance video stream is 32-37 fps,which meets the needs of project application.(2)Multi-target pedestrian tracking.This paper presents a real-time multi-target pedestrian tracking algorithm based on feature fusion.In this algorithm,the target tracking is provided by the yolov3 detection algorithm,and then the trajectory is updated by Kalman filter and Hungarian matching to achieve multi-target tracking of pedestrians.In this paper,we study the occlusion processing method which combines the hog feature,the motion feature and the spatial position information,and optimize the data association effect by combining the distance weight and the feature similarity weight.Using PETS,MOT datasets and practical application scenarios to test and optimize the algorithm,a multi-target tracking algorithm which meets the real-time requirements and has good tracking effect is realized.(3)Design and implementation of intelligent video monitoring system.Applying the research results of multi-target pedestrian tracking algorithm to engineering practice,an intelligent video monitoring system is designed and implemented,which includes the functions of regional intrusion early warning,fast running early warning and pedestrian capture.In order to increase the sharing and interaction of the system,the main functions of the analysis module of the system are displayed by docking with the web module.The main functions of the system are displayed in real time on the front page through Kafka technology,FastDFS technology and RTMP live broadcast technology.The system has a good application effect after the actual scene test.