Research

This study proposes a multi-stage analytical framework linking driving pattern, risky behavior, and road segments to identify potential safety risks in autonomous driving environments where accident data are limited. High-resolution vehicle sensor data were integrated with HD maps to construct a refined time-series dataset. Driving contexts were identified using a Transformer-based deep clustering approach, and potential risk events were detected through a conservative outlier detection method. The detected events were aggregated to generate road-level risk maps, and the influence of geometric road structures on risk formation was analyzed using XGBoost and SHAP. The results demonstrate that the proposed framework effectively captures nonlinear driving patterns and spatial risk characteristics, providing a proactive approach to autonomous driving safety management.