Intelligent Traffic Video Monitoring System Based On YOLOv3 And Multi-target Tracking

At the end of 2018,the total mileage of expressways built and under construction in Guangxi exceeded 7,800 kilometers,with an high-definition video monitoring system installed every 10 kilometers on average to obtain real-time traffic information.However,less than one third of these monitoring systems have the functions of automatic road information acquisition and vehicle abnormal behavior detection.A large amount of monitoring video needs to be manually analyzed,and the supervision efficiency is low,which is not conducive to guiding traffic flow,standardizing vehicle driving behavior and reducing traffic accidents.Aiming at the above problems,this paper designs an intelligent traffic video monitoring system based on deep learning target detection method and multi-target visual tracking method.The system can detect the position,identify the type and track the multi-target trajectory of the traffic target on the expressway,and has the functions of multi-model flow statistics,vehicle abnormal behavior detection and pedestrian snapshot.The main research contents are as follows:(1)Traffic target detection.The principle of three main target detection models(YOLOv1,YOLOv2 and YOLOv3)and the moving target detection method based on Gaussian mixture model are studied and verified by experiments.Experimental results show that the YOLOv3 detection model has the highest accuracy of positioning and type identification for vehicles,trucks,buses and pedestrians in the traffic target,reaching 80%,which is 40% higher than the detection method based on Gaussian mixture model.(2)Single target visual tracking.Firstly,the classification and basic idea of single target tracking method are introduced.Then the principle of Camshift and DSST single target tracking method is analyzed.Finally,the performance of four classical single target tracking methods,Camshift,DSST,KCF and TLD,is verified by experiments.Experiments show that the DSST algorithm can process 101 frames of images per second with a tracking accuracy of 90%.(3)Multi-target visual tracking.A multi-target tracking method based on DSST tracker and multi-target matching is proposed.Firstly,the principle of multi-target matching method based on nearest neighbor and target feature similarity is analyzed.Then,the methods of new target recognition,tracker matching update and target departure discrimination are studied.Experimental results show that this method can track thetrajectory of multiple targets simultaneously.(4)Design of video intelligent traffic monitoring system.The system mainly includes three parts: flow statistics of multiple models,abnormal behavior detection of vehicles,and pedestrian snapshot.In the actual scenario,the data set is established to train the YOLOv3 model.Using YOLOv3 to detect the car,passenger car,truck 3 targets,the use of multi-target trajectory tracking for multi-model flow statistics,traffic target low speed and parking two kinds of abnormal behavior detection,pedestrian snapshot.The results show that the average accuracy of the system is 98.3%.The accuracy of motion direction detection,low-speed motion detection,parking detection,and pedestrian snapshot are all over 75%.The FPS value of the system fluctuates between 17 and 30,basically meeting the real-time requirements.