Research On Test And Evaluation Of Autonomous Vehicles Based On Natural Traffic Flow Generation

Autonomous vehicles have swept the world with their strong posture and become one of the most promising technologies,but autonomous vehicles need to undergo a lot of testing before they can actually go into production to meet commercial demand.At present,the test efficiency of autonomous vehicles needs to be improved,the test scenarios are not sufficient,and the evaluation system still needs to be improved.Therefore,it is particularly important to study the testing and evaluation of autonomous vehicles.This paper relies on the national key research and development program project "Research on the Construction and Simulation Test Technology of Hardware in the Loop Test Environment for Autonomous Electric Vehicles"（No.2018YFB0105103）,taking autonomous vehicles as the research object,from the test scene generation method,test scene data,comprehensive evaluation of several aspects of the research,the specific research content is as follows:（1）Based on the scripted design,the automatic generation and batch testing of autonomous vehicle test scenarios are realized,which improves the test efficiency.Firstly,the scene elements that make up the test scene are classified so that the generated scene covers the test scene of the Autonomous vehicle as comprehensively as possible;secondly,the shortcomings of the traditional manual construction test scene are analyzed;then,according to the degree of abstraction of the scene,the scene is divided into functional scenes,logical scenes and specific scenes,and the three typical scenes of vehicle-motor vehicle,vehicle-non-motor vehicle,and vehicle-pedestrian are analyzed;on this basis,the test scene is built through the script to realize the automatic generation of the test scene;finally,Obtain the result parameters through the script to achieve batch testing.（2）Based on the existing natural driving data set,the improved depth convolution generation adversarial network generates a large number of traffic flow scenarios that meet the characteristics of natural driving,which enriches the test scenarios.Firstly,the appropriate natural driving data set is selected and the data set is preprocessed;secondly,the deep convolutional adversarial generation network and the improved adversarial generation network are used to learn the driving characteristics of the vehicle in the natural driving data,thereby generate a massive traffic flow scene;then,the vehicle’s indicators are evaluated using appropriate statistics,and the results show that the vehicle generated by the improved adversarial generation network is closer to the driving characteristics of the original vehicle;finally,from a macroscopic point of view,The traffic flow scenarios generated by improved adversarial generation network are evaluated,and the results show that the generated traffic flow scenarios are consistent with the natural driving characteristics.（3）Combined with the existing comprehensive evaluation process,the improved weight determination method and evaluation method are used to evaluate autonomous vehicles,and the comprehensive evaluation system is improved.Firstly,the evaluation index system of autonomous vehicles is determined according to the principle of evaluation index selection;then,the method of determining the weight of the index commonly used at present is analyzed,and the weight of the indicator is determined by using the analytic hierarchy method and the improved Critic empowerment method;finally,the fuzzy comprehensive evaluation method is improved by analyzing and comparing the existing evaluation methods to achieve the comprehensive evaluation of autonomous vehicles.（4）This paper builds a traffic flow scenario that conforms to the natural driving characteristics generated by the network through scripts,and uses the improved index weight determination method and the improved evaluation method to carry out the simulation test and evaluation of autonomous vehicles.First,the test scene is built through the script in the simulation software,the test results are extracted and the test results are processed by the script,and the numerical value of the evaluation index system is obtained;secondly,the indicator weight is determined by the analytic hierarchy method and the improved Critic empowerment method,and then the improved fuzzy comprehensive evaluation method is used to evaluate the test vehicle,and finally,the comprehensive score of the test vehicle is determined.Based on the test and evaluation research of autonomous vehicles generated by traffic flow,this paper can improve the test efficiency,enrich test scenarios,and improve the comprehensive evaluation system to a certain extent,which has certain reference value.