Identification of Test Scenarios for Autonomous Vehicles Using Fatal Accident Data

The growing interest from automakers and ride-hailing companies has increased the investment for high automation levels in vehicles. An important challenge in introducing autonomous vehicle (AV) technology to the market is the effort required in the validation. The research shows that AVs have to be test-driven hundreds of millions of miles to demonstrate reliability, which could take hundreds of years. Therefore, the identification of critical test scenarios and reduction of scenario sample space are urgent requirements for providing safe and reliable AVs in a time- and cost-efficient manner.

This article proposes an AV test scenario generation system that creates abstract test scenarios using historical fatal accident data. The method processes and prunes the extensive fatal accident data to generate core test scenarios targeting the reasoning systems of AVs. First, the human-specific factors and the redundant scenario components are filtered out from the crash data so that the accidents can be grouped as core abstract scenarios. The pruned scenarios are then prioritized by severity levels according to the fatality ratio and a relative scaling factor. The functionality of the system is demonstrated by using accident data from the National Highway Transportation Safety Administration (NHTSA). The system reduces the sample space of the utilized dataset substantially, which improves the efficiency of the validation effort. This focused strategy will accelerate the identification of faults in AV systems by complementing the current testing methods.