Research On The Judgment Method Of Vehicle Disobedience And Pedestrian Violation Based On Deep Learning

With the rapid development of China’s economy,more and more people want to have a safe and civilized traffic environment.The traditional traffic monitoring facilities can not meet the increasingly complex traffic environment.More effective supervision means are needed to help the traffic control department ensure people’s traffic safety according to law.Therefore,this paper uses the target detection technology in deep learning to identify the pedestrians and vehicles near the zebra crossing,uses the target tracking technology to track the identified pedestrians and vehicles,and uses Open CV to logically judge the position relationship between the target track tracked in the video image and the zebra crossing and lane crossing according to the Road Traffic Safety Law of the people’s Republic of China.Find out the vehicles that do not yield to pedestrians and take photos to collect evidence.In order to ensure the stability and accuracy of the algorithm and reduce the amount of computation of the algorithm,this paper uses YOLOv5 s with small network structure and less parameters as the detector to identify pedestrians and vehicles near the zebra crossing.Aiming at the problems of low accuracy and unstable model training when the original YOLOv5 s target recognition network is applied to this application scenario,this paper proposes four improvement strategies.Firstly,according to the application scenario of this paper,we build a self built data set,and use K-means clustering algorithm combined with PSO algorithm to generate the Anchor box size suitable for this paper.Secondly,to solve the problem of inaccurate vehicle and pedestrian location and unbalanced samples,the CIOU loss function is used to replace the original GIOU loss function.In order to improve the accuracy of overlapping target recognition,DIOU-NMS algorithm is used to replace the original weighted NMS algorithm.Finally,transfer learning method is used to train the YOLOv5 s target recognition network model.The experimental results show that the improved YOLOv5 s target detection network can greatly improve the positioning ability of vehicles and pedestrians and the detection ability of overlapping targets.Based on the target recognition,this paper uses Deepsort multi-objective tracking algorithm as the tracker of vehicle misdemeanor pedestrian violation decision system.Deepsort algorithm is the optimization algorithm of Sort algorithm.Its data association method combines motion and appearance information,which can solve the problem of inaccuracy of target tracking under occlusion or intense motion.The target position information identified by detector is input into the tracking device to determine the motion track information.The decision module uses Open CV correlation function to make a logical decision whether the vehicle is courteous to pedestrians according to the relationship between the track of the tracked target and the location of zebra crossing area and lane.Finally,the illegal vehicles are photographed for evidence collection,and the last captured photos of illegal vehicles are input into the license plate detection module to obtain and save the license plate information.Experiments show that the system designed in this paper can accurately and continuously determine whether the vehicle is not courteous to pedestrians,which can meet the actual application needs.