Driver-personalized risk assessment in car-following: validation of an optical-flow-based model using SHRP 2...

Traffic crashes remain a major cause of injury and mortality worldwide, yet the effectiveness of crash-preventive safety systems is often limited by a "one-size-fits-all" approach that ignores substantial inter-individual differences in driving behavior. To address this, we tailored an existing risk feeling formula (RFF) to individual drivers' braking styles. Using a subset of the SHRP 2 Naturalistic Driving Study dataset, we extracted 2,911 car-following sequences (gas-off to brake-on). To rigorously prevent data leakage, sequences were split at the trip level into independent calibration and validation sets. For the 38 drivers meeting a minimum sample criterion, the driver-specific RFF model (RFF_prsn) normalized inverse time headway (1/THW) and inverse time-to-collision (1/TTC) by each driver's median values, calculated exclusively from their calibration set. Across the 38 drivers who met the minimum calibration criterion, personalized brake-on medians spanned 0.437-0.974 s^-1 for 1/THW_prsn and 0.049-0.131 s^-1 for 1/TTC_prsn, indicating substantial between-driver differences in preferred following and closing kinematics. Applied to the unseen validation set, RFF_prsn discriminated gas-off from brake-on substantially better than the fixed-weight RFF (area under the curve = 0.942 vs. 0.820). However, the operating-point sensitivity under naturalistic conditions was lower than that observed in our earlier controlled simulator study. Complementary comparisons showed that, in gas-off versus brake-on state discrimination in low- to no-crash-risk situations, a driver-normalized 1/TTC-only model achieved the highest pooled discrimination, indicating that looming (1/TTC) accounts for most of the discriminative information. These results support static personalized normalization as a practical mechanism for aligning an optical-flow-based risk-feeling score with driver-specific braking timing. Future advanced driver assistance system (ADAS)-oriented applications would likely need dynamic, context-aware recalibration to address environmental factors and intra-individual variability.