Research On Construction Of Cut-in Scenario Library Based On Naturalistic Driving Data

As the level of autonomous driving of intelligent vehicles rises,the complexity of autonomous driving systems further increases,and traditional testing methods cannot meet the validation requirements of intelligent vehicle.Scenario-based virtual simulation testing has become an important part of intelligent vehicle testing because of its flexible test scenario configuration,high testing efficiency,high repeatability,and safe and low cost testing process.The test scenario library is the core and source of virtual simulation testing,and whether the constructed scenarios are reasonable or not will directly affect the final test evaluation results of intelligent vehicles.Based on the scenario construction section of the project "Research and Demonstration Application of Intelligent and Connected New Energy Vehicle Testing and Evaluation System",the paper takes cut-in scenario that is the typical potentially dangerous driving scenario as the target scenario and conducts research on the construction method of virtual test scenario library.The main work contents and research results of this paper are as follows.Natural driving data was collected in Changchun.To ensure the effectiveness of the real vehicle natural driving data collection test,the elements of the scenario to be collected and the test route were analyzed and determined,and 12 non-professional drivers with different genders,ages,occupations and driving experiences were recruited according to the real distribution of drivers in China in 2019.The test vehicle was Volkswagen Golf MK8,on which the natural driving data of Changchun City were collected with the addition of a driving recorder,forward millimeter wave radar and VSDS,and the vehicle motion state data were filtered to remove the white noise of the data.Four types of scenario elements significantly related to the danger level of cut-in scenario were obtained.A combination of threshold method and manual video verification was used to extract 59 cases of cut-in scenario data.The risk perception(RP)parameter was selected to characterize the risk level of cut-in scenario,and the parameters significantly related to RP were analyzed by one-way ANOVA and Pearson correlation test in the categorical scenario elements and continuous scenario elements,respectively,and finally confirmed that the ego vehicle longitudinal speed,relative longitudinal speed,longitudinal distance and vehicle cut-in duration were significantly related to the risk level of cut-in scenario.Thus,cut-in scenario is reduced to 4dimensions.Four types of typical cut-in scenarios were obtained for urban road environments.In order to eliminate the effect of different dimension and orders of magnitude on the clustering analysis,the four types of scenario elements were standardized using the Zscore standardization method.The effects of different K-values and initial clustering centers on the clustering effect in the common K-mean clustering algorithm are analyzed.In order to ensure the clustering effect of the scenario data,the K-means clustering algorithm based on density peak point optimization and the K-means clustering algorithm based on hierarchical clustering optimization were used to analyze the cut-in scenario data,and the comparison found that the optimized clustering algorithm improves the clustering effect compared with the common K-means algorithm,in which the clustering method based on hierarchical clustering optimization obtained the largest silhouette coefficient.Therefore,the clustering results obtained by this method were selected as the classification results of the cut-in scenario data,and four types of typical cut-in scenarios in urban road environment were obtained.A virtual testing scenario library for cut-in was constructed and validated.Based on Pre Scan software and the results of cluster analysis,a library of cut-in scenarios was built.To verify its rationality,a joint simulation platform of Pre Scan and MATLAB/Simulink was built,and the mature ACC algorithm was integrated into the four types of cut-in scenarios for virtual validation tests.It is verified that all the four types of cut scenarios can respond to the lane change behavior of the front vehicle in a timely and effective manner and meet the validity requirements of the simulation testing scenarios.