Test case sampling optimization for safety validation of automated driving systems

Area of Science:

• Automotive Engineering
• Artificial Intelligence
• Transportation Safety

Background:

• Automated driving systems (ADS) require rigorous validation using complex test cases mirroring real-world driving.
• Challenges in ADS validation include the complexity of driving environments and the infrequent occurrence of safety-critical events.
• Existing validation frameworks struggle to efficiently capture the full spectrum of driving scenarios, especially rare but critical ones.

Purpose of the Study:

• To develop and demonstrate a novel sampling method for selecting representative and comprehensive test cases for ADS validation.
• To address the challenges of complexity and rarity in real-world driving data for effective ADS testing.
• To enable robust safety validation and performance comparison of ADS against human driving.

Main Methods:

• Introduction of the Kernel Test Case Sampling (KTCS) method.
• KTCS criteria: representativeness (alignment with real-world scenarios) and coverage (capturing high-risk corner cases).
• Application of KTCS to a large-scale naturalistic driving study dataset.

Main Results:

• The KTCS method effectively selects a limited set of test cases that capture long-tailed, rare scenarios.
• The selected cases approximate the overall distribution of naturalistic driving conditions.
• The framework supports accurate accident-rate estimation for fair system comparisons.

Conclusions:

• The proposed Kernel Test Case Sampling method provides a standardized and scalable approach for ADS safety validation.
• This method facilitates accelerated development and deployment of ADS.
• It contributes to building public trust and regulatory confidence in automated driving technologies.