Testing scenario generation and selection for autonomous vehicles using an integrated approach based on real-world accident data

The safety and reliability of Autonomous Vehicles (AVs) are a core concern, which should be validated before application. The critical testing scenarios extracted from historical accidents of AVs can help achieve the efficient safety and reliability testing of AVs. This paper presents an integrated approach that combines a data-driven method with a Bayesian Network (BN). The information including states, states' occurrence likelihoods and quantitative relationships of variables related to scenarios are learned from an AV accident database of California Department of Motor Vehicles (DMV), which is applied to establish a BN. Then, the scenarios are generated and assessed with the BN and a severity matrix. The testing scenarios are selected based on their weighted consequence severity and risk. In this way, this work achieved critical testing scenarios for the Automated Driving Systems (ADSs) and Perception Systems (PSs) of AVs based on the AV accident database.