Driver monitoring fails: 75% of behavior is driver-specific, not complexity-driven

A new arXiv preprint (2607.16847) by Lukas Köning, Nataša Miličić, and Klaus Bogenberger tackles the inverse inference problem of estimating traffic complexity from driver behavior. The team screened 175 behavioral features across five domains—gaze, head pose, longitudinal control, guiding fixation, and scanning strategy—using data from 20 drivers in real urban traffic. Only 31 features exhibited statistically confirmed complexity effects, while 140 were confirmed as null effects through equivalence testing. Mixed-effects variance decomposition revealed a stark imbalance: complexity explains only 1.5% of behavioral variance, driver identity accounts for 23%, and residual variance (unexplained factors) dominates at 75%.

This unfavorable variance ratio explains why all eight feature-level personalization strategies and four classification architectures (e.g., random forest, neural networks) failed under leave-one-subject-out cross-validation, converging at F1 scores around 0.45—barely above chance. The single bright spot is guiding fixation rate (the frequency a driver fixes on a guiding point ahead), which combines speed-robustness, universality across drivers, and minimal inter-driver variation in complexity sensitivity. The authors define three deployment regimes for complexity-adaptive advanced driver assistance systems (ADAS) and argue that the variance structure itself—not feature engineering or model choice—is the primary bottleneck for complexity estimation from behavioral signals.