Design And Implementation Of Decision-Making System For Traffic Condition Assessment Based On Deep Learning

Nowadays,the urban traffic congestion problem is normalized,which not only hinders the development of urbanization,but also affects people’s travel efficiency.Urban traffic congestion is particularly prominent in the crossroads scenario.Most of the existing traffic light control systems adopt the fixed signal timing,leading to a phenomenon that the different directions at the crossroads with different traffic flows have been assigned a same release time.As a result,the accumulation of vehicles in the direction with large traffic flow incurs a traffic congestion at the crossroads.In this situation a blind traffic command may appear.In order to improve the traffic efficiency of crossroads,the introduction of intelligent traffic light system is necessary.However,considering the development of intelligent traffic light system,most systems basically use a single detector,namely radar or video,to collect traffic data,and less consider the combination of the two detectors.Therefore,this thesis thesisproposes a method to collect traffic data by combining video and radar,and develops a decision-making system for traffic condition evaluation based on deep learning.The main work and innovations of this thesis are as follows:1.It is complex and time-consuming to collect the queue length data of vehicles at the intersection by video.In other words,the position coordinates of all queuing vehicles in a lane should be sorted by size in each acquisition procedure.Then,the difference between the maximum and minimum values of position coordinates can be calculated to obtain the queue length data in an acquisition procedure.After verification,it is concluded that obtaining the queue length data from video with an acquisition operation causes an increase of an average of 3.5 seconds in time spent compared to that without acquisition operation.Therefore,in order to maintain the timeliness of the traffic light system,millimeter-wave radar is introduced to attain the traffic data.Firstly,the vehicle position information of each frame captured by radar is obtained by an edge calculator.Then,the data of each frame is processed by the internal algorithm chip of the edge computing device,so that the vehicle queue length in a period can finally be determined.Through the comparison test between video and radar,it is concluded that the accuracy of obtaining the queue length by radar acquisition is 96%,and the delay can decrease to almost zero.2.Based on four scenes of rural,urban,national highway and highway in Xuchang,this thesis makes data sets having a large number of vehicles,and then uses the YOLOv3 network model to train and test the data sets.However,the accuracy of the test results is poor.In order to improve the detection accuracy of YOLOv3 network model,this thesis proposes the I＿YOLOv3 algorithm,which replaces the evaluation index and improves the loss function on the basis of YOLOv3.GIOU is introduced as the evaluation index and the loss function to replace the original evaluation index IOU and the loss function MSE.Finally,under the evaluation index of VOC data set AP-50,the test result of the proposed algorithm shows that the average accuracy of the modified network is improved by 12.8% compared with the original network.3.A traffic condition assessment and decision-making system based on deep learning is proposed.The inputs of the proposed system are the vehicle queue length data collected by radar in the current period Utilizing the DDPG algorithm as the core of the evaluation decision module,the system processes and evaluates the input data through the internal actor network,and the outputs the traffic timing scheme of next period.At the same time,after the traffic data such as the number of queuing vehicles and the average waiting time collected by the video are processed by the weighted formula,the processed results are input into the evaluation decision-making system as a reward value to make the system self-learning,namely,to learn whether the timing scheme of the next period is reasonable.The experimental test shows that the average delay of the proposed system is 0.0774 s.Compared with the fixed timing system,the number of queuing vehicles,the average waiting time,the average delay,the throughput and the queue length of the proposed system are increased by 17.02%,33.33 %,18.18 %,32.29 % and 15.33 %,respectively.