Research On Recognition Technology Of Traffic Violations Based On Deep Learning

In recent years,the traffic flow of urban roads has become increasingly saturated,and the illegal behaviors of vehicles has brought serious burdens to urban traffic supervision.Video intelligent monitoring technology can improve urban traffic management,reduce the cost of urban traffic supervision,and improve the timeliness and accuracy of identifying illegal acts.Based on this background,this research is conducted on the automatic identification technology of illegal behaviors in traffic scenarios.Using the related methods of deep learning,through traffic target detection technology,traffic target tracking technology,and vehicle illegal behavior identification technology,a system scheme for traffic violation identification is designed.The main work is as follows:Aiming at the low performance of vehicle target recognition and localization in traffic scenes,the object detection model of the traffic scenes is studied.The traffic video data was collected on the spot and the target detection dataset Car-Detection of the urban traffic monitoring scenes are produced.Based on the YOLO v4 target detection model,a traffic target detection model is trained in the Car-Detection training set by using the training strategy of setting priori anchors and transfer learning.Experiments show that the m AP of this detection model reaches 99.01%,and the detection speed can reach 28.5 frames per second.For the night traffic detection scenes,the preprocessing method of image contrast and brightness adjustment is introduced.the m AP of the detection model in night validation set can be increased by 4.49%to 92.84%.The experimental results show that the Car-Detection collected in this paper is an effective dataset,and the detection algorithm based on YOLO v4 proposed in this paper is an effective traffic target detection algorithm as well.Aiming at the problem of low matching accuracy of vehicle target tracking under traffic monitoring,the vehicle target tracking algorithm in traffic scenes is studied,and an improved tracking algorithm MF-Deep SORT based on Deep SORT is proposed.MF-Deep SORT introduces Histogram of Oriented Gradient(HOG)feature to represent the target information,which can improve the tracking and matching accuracy.At the same time,the algorithm uses the intersection over union(IOU)to measure the spatial location information,which can improve the tracking accuracy.The traffic multi-object tracking dataset Car-MOT is constructed.Experiments show that in Car-MOT,the MOTA reaches 75.763%,IDsw occurs only twice,indicating that the algorithm has good traffic multi-target tracking performance,which demonstrates the effectiveness of the improved MF-Deep SORT method.MF-Deep SORT algorithm can adapt to different monitoring scenarios,ensure real-time and accurate tracking of vehicle targets,and obtain a complete vehicle trajectory in the monitoring area which is an efficient traffic target tracking algorithm.The identification method of illegal behaviors such as illegal lane changing and driving on the line under video surveillance is studied.Based on the preprocessing of the tracking area and the fitting of the road boundary by least squares method,the mathematical modeling of illegal behaviors of illegal lane changing and lane driving are carried out,and a recognition algorithm Car-Violation for such traffic illegal behaviors is proposed.The relative position relationship between the vehicle target trajectory and the fitting curve is used to determine whether an illegal act occurs.Experiments show that the proposed algorithm Car-Violation can accurately and reliably identify relevant illegal behaviors of traffic vehicles under surveillance.Design and implement a complete set of software and hardware systems for traffic violation recognition technology based on deep learning.The system includes hardware modules such as surveillance cameras,system servers,and software modules such as traffic target detection module,traffic target tracking module and illegal behavior identification module.The actual operation shows that the system can accurately identify illegal acts and collect illegal information.In the test of surveillance video clips of different scenarios,the overall recognition success rate of traffic violations reaches 88%,and the system has good human-computer interaction and real-time performance,which is suitable for actual traffic monitoring.