Generation and Validation of Representative Pre-Crash Scenarios

Doctoral thesis, 2026

Driving automation systems (DAS), including Advanced Driver Assistance Systems (ADAS) and Automated Driving Systems (ADS), are expected to substantially improve traffic safety. Virtual safety assessment is the primary approach for quantitatively evaluating the prospective safety impacts of these systems, but its validity depends critically on the availability of comprehensive and representative pre-crash scenarios. Existing real-world data are limited in quantity and coverage and often suffer from sampling bias, making the generation of synthetic pre-crash scenarios necessary. However, current generation approaches face challenges such as biased or incomplete data and difficulties in validation. In particular, the absence of systematic methods for validating the representativeness of the synthetic scenarios remains a critical knowledge gap.

To address these challenges, this thesis develops an integrated methodological framework for generating and validating representative synthetic pre‑crash scenarios for (prospective) safety impact assessment (SIA) of DAS. The framework consists of two complementary components: 1) a novel approach for generating representative synthetic pre-crash scenarios, and 2) an assessment‑oriented framework for validating their representativeness.

Papers I and II present the proposed scenario generation approach that combines heterogeneous empirical data through model-based parameterization and weighting to construct reference pre-crash datasets. Synthetic scenarios are generated using parametric multivariate models and reweighted to match the reference distributions. The underlying generation logic can, in principle, be applied to conflict-based scenarios with or without collision, but the empirical implementation focuses on rear-end pre-crash scenarios with purely longitudinal dynamics, reflecting current limitations in available datasets.

To address the validation gap, Papers III and IV introduce a Bayesian Region of Practical Equivalence (ROPE)-based framework to assess whether synthetic pre-crash scenarios are practically equivalent to their real-world counterparts for SIA purposes. The framework emphasizes assessment-relevant metric selection, interpretable statistics, and explicitly defined equivalence criteria, and provides diagnostic insight into the sources and implications of non-equivalence.

Overall, the thesis contributes a transparent, reproducible methodology for generating representative synthetic rear-end pre-crash scenarios and a general, assessment-oriented framework for validating scenario representativeness, supporting more accurate and credible SIAs of DAS.