Research On Vehicle Volume Statistics Method Of Two-way Lanes Based On Traffic Monitoring Video

Vehicle volume statistics are the basis for tasks such as alleviating traffic pressure and intelligently managing traffic.It is also a way to use surveillance video to obtain traffic information.It has practical research significance and scenario application value.This paper takes surveillance video as the research object,takes two-way lane traffic statistics as the research task,and realizes the purpose of traffic statistics through three steps of preliminary design,optimization and improvement,and integration of functions.This method can learn the vehicle information of specific road sections,provide diversion decisions for traffic management,prevent congestion,and reduce unnecessary labor expenses.The work content is as follows:(1)Design and implement vehicle detection functionsIn the first step,the UA-DETRAC data set and the self-made data set are merged to form the data set of this article,and YOLOv5 s,which has the smallest network depth and width among the four versions of YOLOv5,is selected for the initial model training.Based on the idea of not making major changes to the lightweight network,it maintains high-speed detection while improving detection accuracy.The second step is to first determine that the overall idea of improvement is to introduce SE into the fusion area of the three major modules of Backbone,Neck and Head in YOLOv5 s.Among them,the improvement plan for Backbone needs to be determined by studying the three different embedding methods of After,Before and Add of the residual block of SE and Bottleneck,and the Add method is proved to be the best choice through experiments.After comparing the detection performance of the three improved networks with the original network,it was decided to use Backbone as the improvement focus area.Finally,the network is optimized according to the idea of deleting the complex and simplifying,removing the improved content in the SE＿Bottleneck CSP1＿1 structure to form SEA1＿Add＿YOLOv5s.A indicates that the improvement position is Backbone,1 indicates that the improvement in the first structure is removed,and Add is the embedding method,referred to as SE＿YOLOv5s.This network is the best solution for improvement.Compared with YOLOv5 s,it can not only obtain a higher m AP than the YOLOv5 s model of 0.014,but also maintain a detection speed of 24 FPS.The third step is to introduce the Focal Loss focus loss function into the output of SE＿YOLOv5s to form SES＿YOLOv5s.This function reconstructs the cross-entropy loss function by adjusting the two parameters α and γ to make the training process pay more attention to positive samples and difficult-to-classify samples,and the detection accuracy Increased and the overall trend of the total loss function becomes smaller.(2)Integrated tracking and counting function based on detectionUse the Deep-Sort algorithm to track the vehicle and extract the trajectory S of the center point of the vehicle,and then delineate the position of the detection line L.If the trajectory S intersects the detection line L,the total counter is increased by one;if the ordinate of the vehicle center point in the current frame is smaller than the previous frame,the up counter is increased by 1,and the down counter is increased by 1.The counting method based on the two detection algorithms can not meet the real-time counting premise.The counting method using SES＿YOLOv5s detection completes the task with an accuracy of 94%,which is 4% higher than the counting method using YOLOv5s.