Applicability of Simulation-Based Road Safety Evaluation Method for Mixed Traffic Conditions Using Real-World Autonomous Driving Data
The purpose of this study is to verify the applicability of a simulation-based road safety evaluation method for mixed-traffic environments using real-world autonomous driving data, and to propose directions for methodological improvement. Technical issues associated with existing simulation-based methodologies include data structure, trajectory connection, segment matching, sample size imbalance, and the definition of reference segments. Analysis data were constructed for 194 road segments using the Lyft Level 5 dataset, and evaluation metrics capturing longitudinal and lateral behaviors and vehicle interactions were derived. Safety Penalty (SP) values were computed for each metric and segment-level relative risk was assessed using an integrated risk score (IRS). Entropy-based weights reflecting the information content of SP distributions were applied and compared using an equal-weighting approach. Based on the equally weighted IRS, high-risk sections were primarily found near intersections, curves, and access points where deceleration, steering changes, and variations in vehicle spacing were prominent. Although applying entropy weights resulted in some changes in the rankings, the relative risk levels of the major high-risk sections were generally maintained. This study identified key issues in the process of applying road safety evaluations using real-world autonomous driving data, including data processing, indicator composition, relative evaluation limitations based on reference sections, and IRS calculation process. Furthermore, considering that IRS is a relative risk indicator rather than an absolute crash risk, this study suggests that additional validation through linkage with traffic crash data is necessary in the future.