Estimation And Simulation Design Of The Number Of Queuing Vehicles Based On Generative Adversarial Network

With the increasing traffic demand,developing intelligent transportation is a feasible solution.Timely and accurately acquisition of current or future traffic situation is of great significance to traffic management and control,which is also an important basis for intelligent transportation.In urban traffic,compared with other indicators,the number of queuing vehicles at intersections can more intuitively represent the traffic state,which can directly act on signal timing optimization.At present,there are many queue length calculation models based on traffic flow theory.But due to the complexity and variability of traffic environment,it is difficult to determine the spatial-temporal relationship and the real distribution of traffic flow,which makes such model-driven traditional methods difficult to model.In addition,the widely used SCATS system in China can't provide the vehicle queue length,it requires a lot of time for timing personnel to count,which is inefficient.Therefore,it is necessary to find a new method to estimate the number of queuing vehicles.In view of the problem of complex calculation and difficult modeling for the model-driven queue length estimation methods,a data-driven model for estimating the lane-level number of queuing vehicles is proposed in this thesis.To avoid the problem that detector data is easy to be missing,this model adopts speed data with better integrity,it uses excellent learning ability and fitting ability of deep learning model to mine the internal law between the average speed and the number of queuing vehicles,and the estimation and classification from the former to the latter is realized.The main work is as follows:(1)The shortcomings of the traditional queue length calculation models and other current estimation algorithms are summarized and analyzed,then the vehiclerunning rules and queuing condition at intersection are studied,which lay a foundation for the subsequent work of this thesis.(2)The feasibility of using speed to estimate the number of queuing vehicles is analyzed.Firstly,the quality of SCTAS detector in Hangzhou is statistically analyzed,which proves that the traditional model based on detector data can't be widely used.Then collecting the variation trend of average speed and average flow of a road within24 hours,and analyzing the relationship between the speed and the number of queuing vehicles in one day.Finally,the feasibility of the idea is verified by using LSTM network and simulation data obtained from SUMO,which determines the direction for subsequent algorithm proposed in this thesis.(3)A model for estimating the number of queuing vehicles based on improved ACGAN is proposed.which combines the characteristics of Bi-directional Long Short-Term Memory(Bi-LSTM)network and auxiliary classifier generation adversarial network(ACGAN).According to the implicit relationship between the speed change sequence and the number of queuing vehicles,the generator uses Bi-LSTM to capture the time correlation between them and generates the minimum and maximum number of queuing vehicles.The discriminator comes from ACGAN,which can realize multi-classification from the number of queuing vehicles to the congestion level labels while distinguishing the true and false samples,so as to better represent lane's congestion situation.In order to avoid the problem of gradient disappearance,the initial true-false binary classification tasks in discriminator is abandoned,and the Wasserstein divergence is introduced instead of JS divergence to measure the distribution distance between the true and false sequences with the corresponding objective function is modified.In addition,a gradient prediction step is added into network update process,so as to ensure the solution path has a more stable tendency towards the saddle point.Experimental results show that the model has good stability and convergence speed,and the generator obtains the minimum estimation error on more than half of the lanes while the discriminator's classification accuracy reaches 93.29%.The estimation and classification performance of this model are better than other comparison algorithms,which can meet most estimated demands of the number of queuing vehicles.(4)A data-driven vehicle queue micro-simulation framework is designed.Elaborating the structure of the framework and the function of each module and applying the estimation algorithm in this thesis to this framework.On this basis,the simulation idea of vehicle queue and the simulation process are further studied.Finally,the main functions of this framework are implemented.The designed simulation framework can make up the shortcomings of the SCATS system,so as to improve the situation that timing personnel check intersection monitoring manually to obtain the number of queuing vehicles and enhance the efficiency of signal optimization.