Critical scenario sampling method for autonomous vehicle test based on the concept of scenario space partitioning

Abstract

Sampling critical scenarios from the natural driving datasets (NDD) to build a test scenario library has proven to be one of the most effective approaches. Critical scenarios typically follow a long-tail distribution with the “class imbalance” characteristic, posing a challenge in balancing between exploitation and exploration to ensure the efficiency and comprehensiveness of the test scenario library simultaneously. Researchers employed the ϵ-greedy policy to extract risky scenarios and achieved encouraging progress in ensuring testing efficiency. However, due to the ε-probability random sampling, critical scenarios with low probability may be missed from the test scenario library, limiting the test comprehensiveness. To bridge this gap, this paper proposes the Scenario Space Partitioning-Critical Scenario Sampling Method (SSP-CSSM) to efficiently sample critical scenarios from a long-tail distribution. Firstly, the concept of Scenario Space Partitioning (SSP) is proposed to decouple exploitation and exploration as sampling the risky scenarios and the rare scenarios respectively. Next, the normalized time to collision (NTTC) and normalized Shannon Information (NSI) are designed to represent the risk and rarity of a scenario. The intersection of NTTC and NSI projections is defined as the boundary to divide the scenario space into a high-risk sub-region and a high-rarity sub-region. Finally, different sampling methods and weight functions are proposed for each sub-region to efficiently sample critical scenarios. In the high-risk subregion, the NTTC is regarded as sampling weight to sample risky scenarios for exploitation, achieving an equivalent sampling efficiency as the ϵ-greedy sampling policy. In the high-rarity region, the sampling weight function based on $\mathrm{NSI}$ is designed to amplify the granularity of rare scenarios for exploration, capturing critical scenarios missed by the ϵ-greedy policy. Simulation results showed that the SSP-CSSM method sampled a significantly larger number of crash scenarios, thereby improving the efficiency and enhancing the comprehensiveness of the test scenario library.