A Versatile Approach to Evaluating and Testing Automated Vehicles based on Kernel Methods

The auto companies have been competing to get their automated vehicles (AVs) ready on road for years, yet there is still none available in the market. Partly, this is due to the challenging task of robustly testing and guaranteeing the safety of an AV before its release. Companies have been trying different methods such as road test [1], [2], computer simulation test [3] and human-vehicle interaction test [4], [5], yet providing safety certificate for an AV system is still open for solving [1]. Assisting the endeavors of solving this problem, the U.S Department of Transportation has released a new AV policy: A Vision for Safety 2.0 [6]. This official document standardizes the required safety features of an autonomous vehicle, providing guidance and clearer pathways for the various stakeholders aiming to certify the safety of their AV systems. However, even with this newly published official guideline, the testing standard remains unclear while the AV target release is quickly approaching. Thus, an effective and efficient testing method for an autonomous vehicle is an urgent need under this background. Traditional vehicle safety tests are based on crash databases collected from crashes or dangerous scenarios, such as the CSD and GIDAS crash databases [7]. However, the information logged in these databases is limited so that it is difficult to reconstruct and analyze the dangerous scenarios.