Research And Application Of Traffic Flow Detection And Intersection Signal Control

With the growing popularity of automobiles,traffic congestion in China is getting increasingly severe,and overcoming the problem of urban congestion has been a focus of intelligent transportation research.In the intelligent transportation system,traffic flow is an important basis for traffic management,and intersection signal control is a concern for people’s livelihood.The development of intersection signal control scheme through real-time detected road status plays an important role in improving traffic efficiency.The main work of this thesis is as follows:(1)This thesis presents a method to detect traffic flow based on traffic video.The ADDeep SORT algorithm solves the problems that the cross-parallel ratio in Deep SORT algorithm has more matching errors and the feature extraction network is not applicable to traffic scenarios.In order to improve the accuracy of vehicle volume detection in the road environment,the YOLOv3 network structure is modified and the anchor box is regenerated.To verify the effectiveness of the AD-Deep SORT algorithm with the YOLOv3 algorithm as the target detector,experiments are conducted to detect the traffic volume in the video.The results show that ADDeep SORT achieves better results compared to the traditional methods.(2)To tackle the difficulty of a single optimization index in the present model,this thesis provides a temporal model for multi-objective optimization of single intersection signals.The model takes real-time traffic flow as input and provides signal timing schemes with optimization targets of junction driving capacity,average vehicle delay length,and average vehicle stopping frequency.To simplify the model,the multi-objective optimization is reduced to a singleobjective optimization,and the model’s optimal solution is found using a genetic algorithm.When compared to the actual road timing scheme and the Webster model,the experimental findings reveal that the model scheme effectively enhances the total throughput efficiency of a single intersection.(3)The MA-PPO road network timing control technique is proposed in this thesis,which is based on deep reinforcement learning.Joint states and joint reward functions are utilized in the agent to solve the problem of a lack of connection between road network intersections.To improve timing flexibility,a technique with no set phase order action space is used.The network structure is modified to respond to the traffic environment,and the feature information of the intersection intelligences is fully extracted.In comparison to standard timing methods and other deep reinforcement learning methods,simulation results show that the MA-PPO algorithm delivers good results in both queue length and delay time performance metrics.