Autonomous driving accelerated evaluation method for independent/dependent variables based on importance sampling

The safety of Autonomous Vehicles (AVs) is crucial to the development of the autonomous driving field. The accelerated evaluation methods based on scenario simulation have become a hot research direction due to low test cost and high test efficiency. Among them, the Importance Sampling (IS) method has attracted much attention. However, the IS method based on variable independence is limited in its application actual scenarios. To address this issue, this paper proposes an accelerated evaluation method compatible with both independent and dependent variables. For independent variables, a non-parametric Kernel Density Estimation (KDE) method is employed for distribution fitting, combined with IS and Bayesian Optimization (BO) to present a non-parametric sampling strategy. For dependent variables, by utilizing the Copula model and Maximum Likelihood Estimation (MLE) to capture the dependencies between variables, the joint distribution of dependent variables can be obtained, facilitating accelerated evaluation in conjunction with IS and BO. Furthermore, this paper selects the relative half-width as the convergence indicator and sets a reasonable threshold according to the probability distribution of variables. Through simulation testing in cut-in scenarios, this accelerated evaluation method not only effectively accommodates both independent and dependent variables but also achieves an increase in testing efficiency of over 200 times compared to Monte Carlo methods. Meanwhile, through the analysis of different variable combinations, it is found that this method can select the variable combination with the highest test efficiency according to the importance distribution of each variable combination, providing new ideas and technical support for the theory of accelerated evaluation.