Environmental and contextual predictors of driver drowsiness and distraction from a naturalistic driving study on...

This study investigated environmental and contextual predictors of driver drowsiness and distraction alerts detected by camera-based driver monitoring systems (DMS) during naturalistic driving, using negative binomial regression to model rare-event count data with environmental, temporal, and traffic covariates.

Data were collected from two public buses equipped with DMS over a 3-month period on the Hyderabad-Kodad National Highway NH-65 corridor in India, yielding 1,546 ten-minute driving segments. Outcome variables were counts of drowsiness alerts and distraction alerts. Predictors included time of day (early morning, morning, afternoon, evening, night), traffic conditions (congested, moderate, free-flow), temperature categories (cool, warm, hot), and elevation zones (low, mid, high) as categorical variables; vehicle speed and trip duration were analyzed as continuous variables. Negative binomial regression models were fitted using progressive model selection, with sensitivity analyses including robust standard errors, outlier exclusion, bootstrap confidence intervals, and 5-fold cross-validation.

Morning driving (06:00-09:59) exhibited 158% higher drowsiness rates than afternoon driving (incidence rate ratio [IRR] = 2.58, 95% confidence interval [CI]: 1.44-4.62, p = .001), while evening driving showed 19% higher distraction rates (IRR = 1.19, p = .021). Free-flowing traffic was associated with 79% lower drowsiness (IRR = 0.21, p = .008) but 45% higher distraction (IRR = 1.45, p = .024), consistent with opposite attentional mechanisms. Low-elevation driving (<200 m) was associated with 7-fold higher drowsiness risk (IRR = 7.09, p = .002), which may reflect monotonous flat highway geometry. Warm/hot temperatures (25-30 °C) were associated with 58% to 61% lower drowsiness (p < .014), whereas hot conditions were associated with 41% higher distraction (p = .011). Trip duration strongly predicted distraction (0.6% increase per minute, p < .001) but not drowsiness. Cross-validation confirmed model stability (cross-validation < 10% for both outcomes).

Environmental and contextual factors were significantly associated with DMS-detected impairment events, with traffic conditions showing opposite associations for drowsiness versus distraction. Findings support development of context-aware DMS algorithms that adapt alert thresholds based on time of day, traffic state, temperature, and trip duration to enhance targeted crash prevention interventions.