Scenario Extraction And Automated Testing Of Pedestrian And Vehicle Based On Naturalistic Driving Data

In road traffic,when a traffic accident occurs between a vehicle and a pedestrian,it is usually the pedestrian who is injured or,in severe cases,the injured party who is killed.Therefore,the construction and design of pedestrian-vehicle scenarios are very important for the safety testing of smart vehicles.Accident data and naturalistic driving data contain a large number of interaction scenarios between pedestrians and vehicles on the road,which are important sources for constructing a database of pedestrian-vehicle scenarios.The purpose of vehicle safety testing is to detect the boundary of the system and the response of the vehicle under dangerous driving scenarios.Therefore,it is necessary to divide the pedestrian-vehicle scenarios into non-dangerous and dangerous scenarios,and take the dangerous scenarios as the main test scenarios.Based on the national key research and development project(No.2018YFB1600800),this paper has carried out the research on the test and evaluation of the intelligent vehicle road system.Based on naturalistic driving data,this thesis proposes a "data screening-variable extraction-clustering-scenario automated testing " method for the extraction and testing of pedestrian-vehicle scenarios.The main content has the following four aspects:(1)Naturalistic driving data processing and establishment of non-dangerous/dangerous incident data sets for pedestrians and vehicles.This thesis uses the naturalistic driving data set to filter out the data of pedestrians in front of the vehicle based on the information of the front object recognized by the radar in the data.Through the data extraction,invalid and missing value processing,scenario element labeling and pedestrian track calculation,221 valid pedestrian and vehicle incidents were obtained after processing.According to the speed limit and the shortest braking distance of the vehicle driving area,the preliminary extraction of dangerous incidents is carried out,and then the final pedestrian-vehicle incidents are obtained by manual verification.In the 221 pedestrian-vehicle incidents,56 dangerous incidents and 165non-dangerous incidents were extracted.(2)The random forest algorithm is used to extract the variables that affect the danger of people and vehicles.Because the data dimension is high,not every variable will affect the danger of pedestrians and vehicles.Therefore,the importance of variables in the random forest is sorted,and 13 important variables that affect the danger of people and vehicles are screened out,which are pedestrian movement direction,vehicle movement direction,vehicle travel area,relative longitudinal and lateral distance,relative longitudinal and lateral speed,GPS speed,longitudinal acceleration,steering wheel angle,angular velocity and TTC,these variables are used as input to the subsequent algorithm.(3)Extract pedestrian and vehicle scenarios from non-dangerous and dangerous incidents of pedestrian and vehicle.The data of 13 important variables are obtained from the data set of pedestrain and vehicle non-dangerous/dangerous incidents,and the two kinds of pedestrian-vehicle incidents are clustered and analyzed by the systematic clustering algorithm,and 4 types of dangerous scenarios and 6 types of non-dangerous scenarios are obtained.This scenarios are the logical scenarios.(4)Carry out scenario automated testing for dangerous scenarios.Scenario automated testing can speed up vehicle testing from dangerous scenarios to specific scenarios.Based on the idea of an adaptive proxy model,a scenario automated testing method is proposed.The content of the scenario automated test includes: use Latin hypercube sampling in the scenario test matrix to obtain the initial sample points;then use Pre Scan to perform the scenario simulation test to obtain the response values of the initial sample points,and train the proxy model with the initial sample points and response values;the update point and response value are selected to update the model;finally,the test results of proxy models using different classification algorithms are compared with evaluation indicators.The results show that scenario automation testing can greatly reduce the number of tests for specific scenarios.