Abstract
Advanced driving assistance systems (ADASs) provide rich data on vehicles and their surroundings, enabling early detection and warning of driving risks. This study proposes a short-term risk prediction method based on in-vehicle perception data, aiming to support real-time risk identification in ADAS environments. A variable sliding window approach is employed to determine the optimal prediction window lead length and duration. The method incorporates Monte Carlo simulation for threshold calibration, Boruta-based feature selection, and multiple machine learning models, including the light gradient-boosting machine (LGBM), with performance interpretation via SHAP analysis. Validation is conducted using data from 90 real-world driving sessions. Results show that the optimal prediction lead time and window length are 1.6 s and 1.2 s, respectively, with LGBM achieving the best predictive performance. Risk prediction effectiveness is enhanced when integrating information across the human-vehicle-road environment system. Key features influencing prediction include vehicle speed, accelerator operation, braking deceleration, and the reciprocal of time to collision (TTCi). The proposed approach provides an effective solution for short-term risk prediction and offers algorithmic support for future ADAS applications.