Research On Traffic Flow Detection Based On Deep Learning

With the rapid development of the economy,the number of vehicles increases yearly.The problem of traffic congestion has become increasingly severe,and the method of real-time detection of traffic flow has been adopted to divert and control road vehicles,which can effectively alleviate traffic congestion.Many methods were developed for realtime detection,while one of the popular methods is the detection with deep learning,whose principle is that based on deep convolutional neural networks,the targets in the video frames are classified and located,and then the Kalman filter algorithm is utilized to track the targets.Finally,the vehicle flow is counted by virtual loop method.However,due to the complexity of the actual traffic scenario,vehicles in video frames may be inspected falsely or missed due to target size,occlusion,and intensity changes.In addition,it isn’t easy to achieve real-time detection when complex algorithms are deployed on edge computing devices.In view of the above problems,the existing object detection algorithm will be improved in this paper,then combine the Deep SORT multitarget tracking algorithm to realize traffic flow detection,and deploy the algorithm on the domestic edge computing AI chip to achieve real-time detection of traffic flow at the edge end,the specific work content is as follows:(1)To address the problem of low accuracy of small target detection in actual scenarios,a multi-scale feature fusion strategy was adopted to improve the YOLOv5 s target detection algorithm network.The shallow feature map with more small target information is down-sampled and fused with the Neck part to obtain a new multi-scale feature map and predict the small targets.(2)To address the problem of realizing real-time detection on the edge computing AI chips with limited computing resources for complex models,a network structure reparameterization approach for vehicle detection algorithms has been proposed,in which the Rep VGG module was incorporated into the YOLOv5 s network to improve the speed of model inference.(3)To address the problem of low accuracy in target detection for complex scenarios where vehicles are blocked and multiple intensity light environments,the data enhancement method is used to expand the vehicle image data taken on the road to enhance the generalization of the model,and Labelme image labelling tool was used to label the vehicle dataset.Finally,the model was trained,and performance analysis was carried out.(4)The development environment for the vehicle flow detection system was established.The system hardware and software framework were designed,while the Rockchip RV1126 platform development tool was utilized to quantize the model,which results in accelerating the speed of model inference at the edge device,and finally the improved vehicle flow detection algorithm on the edge computing platform RV1126 was deployed.