Test Scenario Generation Method for Intelligent Vehicles Based on Neighbor Object Region Representation and Conditional Variational Autoencoder
Scenario-based virtual testing is a necessary approach to developing intelligent vehicles with high safety and reliability.Automatic scenario generation technology is valuable for constructing the test scenario library for intelligent vehicles.Therefore,a scenario generation method based on neighbor object region representation(NORR)and conditional variational autoencoder(CVAE)was developed for dynamic test scenarios with multivehicle to rapidly generate complex test scenarios and control the types of generated scenarios.First,the NORR method was developed to describe the highway scene situation,and the key information of vehicle objects in the test scenario was converted into grayscale images with uniform sizes.Next,the HighD dataset of naturalistic vehicle trajectories was used to extract many scene fragments,and the real-scene library was constructed after data normalization processing.Based on this,the CVAE-based generative model was trained with the number of vehicle objects in the scene as the conditional parameter,which could generate a dynamic test scenario containing eight vehicle trajectories.By calculating the matching error,coverage degree,and unreasonableness of the generated sample set,the performances of the generative model were analyzed in terms of sample authenticity,diversity,and rationality.The verification results show that ① compared with the random trajectory sampling method and generative adversarial network-based model,the quality of the scenario samples generated using the variational autoencoder model is the best.The average matching error of the generated samples is lower than 5.22,the coverage degree is up to 57.2％,and the proportion of unreasonable samples only accounts for 1.7％.② The proposed NORR method helps improve the scenario generation effect of generative models.③ The CVAE model can establish the correlation between the conditional input and the generated results.By adjusting the conditional parameter,the number of vehicle objects in the generated scene can be varied.
automotive engineeringtest scenario generationCVAEintelligent vehiclegenera-tive model
NETL
NSTL
万方数据
