Research On Traffic Flow Generation Model And Its Rationality Evaluation Based On Improved GAN

In recent years,with breakthroughs in various software and hardware technologies,major manufacturers have joined the research and development of autonomous vehicles,and autonomous driving technology has become the hottest research direction.Studies have shown that an autonomous vehicle needs to travel at least 1 billion miles to verify that it has better performance than human drivers at a 95% confidence level.Obviously,for this higher test requirement,safety verification cannot be completed through a single real vehicle test.Therefore,it is extremely important to use virtual simulation technology to simulate road traffic conditions and verify the safety of the auto drive system.However,using expert experience based traffic flow modeling methods or mechanism based traffic flow modeling methods has problems such as low testing efficiency and few scenarios.Therefore,in order to enrich traffic flow scenarios and improve test efficiency,this article uses confrontation generation network technology to train and learn actual traffic flow characteristics,and ultimately obtains a model that can automatically generate specific traffic flow test scenarios.In order to obtain this generation model,this article conducts research from the following aspects: selecting and processing data sets,evaluating the rationality of the generation model,and selecting and improving training models.The specific research content is as follows::(1)According to the specific needs of generating the model,select the appropriate dataset,and through processing,ultimately obtain a training set that meets the training requirements.Firstly,according to the characteristics of the traffic flow studied in this paper,select a suitable public dataset for natural driving of vehicles;Secondly,select appropriate congestion classification criteria for the public dataset in the selected area,and perform processing on the original dataset such as noise removal,frame completion,prescribed sampling interval,time normalization,and data filling.Refine the original dataset to extract tens of thousands of separate traffic flow samples;Finally,classify and save the feature information of each sample to obtain a training set that meets the training requirements.(2)In order to analyze and evaluate the sample generation effect of the generation model,a comprehensive evaluation method based on Spearman correlation coefficient is used to evaluate its rationality from the perspective of accuracy and diversity.Firstly,select appropriate evaluation indicators based on the characteristics of traffic flow samples;Secondly,compare and analyze the advantages and disadvantages of different indicator weight determination methods.According to the characteristics of evaluation indicators and the relationship between different indicators,select an indicator weight determination method based on Spearman correlation coefficient,and calculate the membership functions of different indicators in each scenario;Finally,the generated model generates a large number of samples under each scenario,calculates each sample’s score based on the membership function and indicator weight,summarizes the score distribution of the generated sample and compares it with the score distribution of the training set samples,analyzes the accuracy of the generated model,and compares the indicator ranges of each sample in the training set and the generated model in the form of a scatter chart,analyzes the diversity of the generated model,and achieves a reasonable evaluation of the generated model.(3)Based on the processed training set and rationality evaluation method,an improved adversarial generation network(GAN)is used to train it,and ultimately a generation model that can automatically generate massive traffic flows on demand is obtained.Firstly,select a basic confrontation generation network for training to find out the problems in training and the reasons for poor training results;Secondly,aiming at the problems of the original adversary generated network: gradient disappearance and mode collapse,the original network is improved by replacing the network structure of the generator and discriminator,customizing the loss function,and selecting the appropriate optimizer.The specific effects of different improvement methods are obtained through training;Finally,by analyzing and comparing the results of different improvement methods,the optimal network structure improvement scheme is obtained: the overall structure uses conditional auxiliary classifiers to generate a network against each other,the loss function uses bulldozer distance and least squares loss,the discriminator uses residual networks,and the generator uses a U-Net based multi generator network.