Research On Traffic Flow Detection Based On Deep Learning And Edge Task Offloading

In order to solve a series of problems caused by traffic congestion,intelligent transportation system emerges as the times require.Through real-time acquisition of urban road traffic flow information,and on this basis,provide intelligent guidance to alleviate traffic pressure and reduce environmental pollution.One of the key technologies is traffic flow detection.In recent years,intelligent video detection method has been widely concerned,but the image algorithm has high complexity and needs to consume a large amount of computing resources.So the traffic flow detection in ITS usually adopts cloud computing mode.All videos are collected and transmitted from the edge of the network to the cloud computing center.The cloud computing center runs image algorithm to detect traffic flow.However,the problems of large storage of traffic surveillance video and limited network bandwidth bring great challenges to achieve efficient traffic flow detection in ITS.In this thesis,traffic flow detection in ITS is taken as the research task.Aiming at the road environment in traffic surveillance video,a method of offloading traffic flow detection task to the edge is designed and implemented by combining deep learning and edge computing technology.Specific research work is as follows:(1)A method for edge unloading of traffic flow detection task is proposed.Cloud Computing Center iterates deep learning vehicle detection and tracking algorithms based on traffic surveillance video,including training and testing of algorithm model.Then the trained model is migrated to the edge intelligent device Jetson TX2,and the traffic flow detection algorithm is further implemented on the device to complete the task of real-time traffic flow detection at the edge.(2)A vehicle detector based on YOLO(You Only Look Once)is designed and implemented.Aiming at the problems of low precision and slow detection speed of traditional object detection algorithm.In this thesis,the YOLO network with high real-time performance is used to detect vehicles,and the network parameters are optimized by simplifying the model,reducing the detection class category and using K-Means clustering method to modify anchors.After training and testing,a vehicle detector with good accuracy and high real-time performance is finally obtained.(3)A multi-target tracking algorithm based on Deep-SORT(Simple Online and Realtime Tracking with a Deep association metric)is implemented.Traffic flow detection often needs to be combined with vehicle detection and tracking.In this thesis,a multi-target tracking algorithm based on Deep-SORT is implemented.To solve the problem that the feature extractor of the algorithm does not learn the vehicle features,we use a vehicle reidentification data set to retrain the Deep-SORT tracking model.After testing,the retrained model has good tracking accuracy and high real-time performance.(4)A multi-target real-time tracking counter based on YOLO and Deep-SORT is designed and implemented.This thesis combines YOLO detector and Deep-SORT tracking algorithm,and uses virtual detection line method to count the number of vehicles passing through the detection line to realize traffic flow detection.Furthermore,an interval multi-frame tracking method is proposed to optimize the counter.After testing,a high accuracy real-time traffic flow detection is realized on the edge intelligent device Jetson TX2.Finally,the traffic detection task is successfully unloaded from the cloud to the edge,which makes the traffic surveillance video be processed at the edge,alleviates the pressure of transmission and storage,and improves the efficiency of traffic flow detection in ITS.