Direction Field-Guided Intelligent Vehicle Testing With the Gradients of the Logarithmic Evaluation Distributions

Although intelligent vehicles have been widely used, the assurance of their safety remains challenging, as evidenced by the high rate of disengagement and traffic accidents. Before the deployment of intelligent vehicles, comprehensive testing and validation are indispensable to ensure the safety of passengers and pedestrians, as well as to drive technological advancements. Existing studies have developed numerous artificial intelligent methods for the testing and validation. However, these methods encounter the challenges regarding the time-consuming expenditure and interpretability of the complex network structures they adopted. Moreover, it is difficult for these methods to design sampling functions that are tailored to different testing objectives. Therefore, this study proposes a new testing method by using the gradient of the logarithmic function of the evaluation distribution as a direction field to generate testing scenarios while making the sampling of testing scenarios towards critical regions. First, the operational design domain (ODD) is segmented into distinct subintervals to decrease the complexity of modeling. In these subintervals, the available traffic states are modeled as numerous high-dimensional Gaussian distribution functions, and the overlapping regions are represented as a Gaussian mixture model, i.e., a linear combination of adjacent Gaussian models. Then, the distribution of evaluation metrics within each interval is assumed differentiable. Taylor expansion is performed to derivate the sampling direction of critical scenarios conditioned on quantitative indexes. And the direction field is approximated using a simple neural network. Afterwards, model parameters within each subinterval are estimated using naturalistic driving data, and critical testing scenarios are generated. Finally, the proposed testing method is verified through an extensive variety of experiments. Results demonstrate that the proposed direction field – guided testing method efficiently uncovers more critical scenarios efficiently and boosts the testing efficiency.