Research On Traffic Flow Detection Based On Traffic Video

The development of cities has brought about a series of urban traffic problems.Intelligent Transportation System(ITS)is considered to be the best way to solve the traffic problems.Traffic flow is an important basic information in ITS.How to detect traffic flow quickly and accurately is an important research content of ITS.With the development of computer vision and Internet of Things technology,traffic surveillance has become a new technology of traffic information collection and will play a greater role,thus,how to detect vehicle and traffic flow quickly and accurately has become a key problem.Traditional vehicle detection methods,such as background subtraction,frame subtraction and optical flow,have some limitations.Object detection based on deep learning has gradually become the mainstream method.In this paper,deep learning is applied for traffic flow detection in traffic video.Firstly,aiming at the lack of annotated data,a method of constructing vehicle detection model based on scene transfer learning is proposed.Then,for the possible situation of vehicle missing detection and false detection,a virtual traffic detection method based on vehicle tracking is proposed.Finally,the topological relationship of road sections is introduced into traffic flow detection,and the traffic flow topological model is established to correct the traffic flow detection results.The main studies and contributions are summarized as follows:(1)The study of construction method of deep learning vehicle detection model based on scene transfer learning: In order to solve the problem of lacking labeled data,on the basis of open data sets,transfer learning and deep learning are combined,instance-based transfer learning and parameter-based transfer learning strategies are adopted,and a vehicle detection model with good performance is constructed quickly.(2)The study of traffic flow detection based on deep learning: Aiming at the vehicle bounding box obtained by the deep learning vehicle detection,considering the possible situation of vehicle missing detection and false detection,and combining the ideas of traditional traffic flow detection method based on virtual detection region and target tracking,a new method of traffic flow detection based on fusion of virtual detection region and vehicle tracking is proposed.(3)The study of missing alarm suppression and false alarm suppression for vehicle missing detection and false detection: For the possible situation of vehicle missing detection and false detection,the missing alarm suppression module based on vehicle tracking and the false alarm suppression module based on bounding box size statistics are designed in the traffic flow detection model to avoid the errors of traffic flow counting caused by vehicle missing detection or false detection,and further improve the accuracy of traffic flow detection.(4)The study of traffic flow correction based on the traffic flow topological model:Considering the topological relationship among the road sections in the road network,the traffic flow topological model is established to correct the traffic flow detection results,and the result of road sections which is unreliable will be corrected through the road sections with reliable results.The experimental results show that the proposed method can quickly build a deep learning traffic flow detection model without label data,and the model has high accuracy of traffic flow detection in urban traffic video.In terms of detection efficiency,this method has high real-time performance,and can meet the practical application requirements,and can be used for real-time traffic flow detection.