Traffic Flow Information Statistics Research And System Implementation Based On Deep Learning And Edge Computing

Traffic flow information statistics,as a basic part of intelligent transportation construction,is a very important task to help people analyze road flow pressure and carry out traffic light control,etc.The current equipment for traffic flow statistics is mainly radar and coil,which is costly and troublesome to maintain.Using four-way surveillance cameras at intersections requires complex operations such as image fusion and object de-duplication,as well as easy to produce occlusion between vehicles in the direction of single-way tilt angle,which makes it difficult to obtain the current road condition information simply and obviously.Therefore,this paper proposes a statistical computation method of traffic flow information based on high level camera and deep learning for the problems of dense vehicle targets and dramatic scale changes,and builds a relevant engineering application system by combining edge computing.The main work of this paper is as follows.(1)In this paper,the cascaded YOLOX target detection model(CAS-YOLOX)is improved based on the YOLOX target detection model.For the problem of dense vehicle distribution,a cascade detector applicable to the single-stage anchor-free approach target detection model is proposed to improve the quality of the detection frames.For the problem of drastic changes in vehicle scales,a layer feature correlation extraction and feature fusion module(CEFF)is proposed to achieve adaptive fusion of adjacent layer features through spatial information adjustment and extraction of layer feature correlation operations to mitigate the effects of drastic changes in target scales.In this paper,a high-resolution branch is used for the small target task,which greatly improves the detection accuracy of the model.The experimental results show that CASYOLOX outperforms many current popular models with good accuracy and robustness.Finally,the effectiveness of each component of CAS-YOLOX is demonstrated by extensive ablation experiments.(2)In this paper,Coordinated Attention Based Multi-Target Tracking Network(CoordSORT)is proposed based on DeepSORT target tracking algorithm for improvement.For the short-time occlusion problem during vehicle tracking,this paper combines the cross-channel domain attention module and CSP structure to propose the cross-channel domain coordinated attention network(CoordECANet),which improves the extraction capability of the apparent features of the target.Experiments show that in the case of short-time occlusion,CoordSORT has less transformation of the same target ID and better tracking performance compared to DeepSORT.This paper proposes a novel traffic flow statistics framework,whose error rate is 7.8%(complex scene)and3.7%(simple scene)tested in the actual captured video.The proposed framework meets the basic requirements of actual monitoring.(3)In response to the current problem of large volume of video data for massive traffic monitoring,direct transmission to the server will occupy a lot of bandwidth and computing resources,this paper proposes a way to deploy the traffic flow information statistics framework through the Tensor RT framework in combination with edge computing to reduce the transmission bandwidth and relieve the computing pressure on the server by transmitting structured data instead of video data through edge devices.(4)This paper builds a traffic flow information statistics system based on the proposed algorithms.In order to visualize the traffic condition of the intersection obtained through each part module of this system,this paper builds a visualization platform through Vue framework and uses Blender software for 3D modeling to understand and analyze the traffic condition by the Digital Twin.The proposed system is simple to deploy,easy to maintain,highly scalable,and can accurately statistic and present the traffic flow.