Improving Driver Drowsiness Detection through Temporal, Contextual, and Hierarchical Modeling.

摘要

Drowsiness-related vehicle crashes are a persistent and substantial hazard on today's roadways. Drowsiness mitigation technology promises to reduce these crashes by detecting drowsiness and providing interventions to drivers. Mitigation technology relies on accurate detection algorithms to inspire driver trust and appropriate use of the technology. This dissertation investigates gaps in the current drowsiness detection literature and iteratively develops a series of temporal, contextual, and hierarchical models to address these gaps. This dissertation uses data collected from a high fidelity driving simulator to predict drowsy-related lane departures. The three studies discussed in this dissertation investigate the effects of dynamic graphical modeling structures, road context integration, and hierarchical context integration on model detection performance. The investigation of dynamic graphical models included Hidden Markov Models, Hidden semi-Markov Models, and Conditional Random Fields. The study of road context integration investigated distributional parameters, Fourier transforms, and Symbolic Aggregate Approximation for generating road context from vehicle speed and acceleration data. The hierarchical context study investigated generation of both road-context and maneuver-level context from speed and acceleration data using Symbolic Aggregate Approximation time-series analysis. The three studies showed a benefit of including temporal dependencies and maneuver-level context in drowsiness detection algorithms. Including both of these factors significantly reduced false positives generated by the algorithm relative to PERCLOS, a commonly applied algorithm in the drowsiness detection literature, and a steering-based algorithm that did not consider temporal or contextual factors. Maneuver-level context increased detection performance relative to both road type and a hierarchical combination of maneuver-level and road type contexts. State duration modeling undermined model performance and was not effective for drowsiness detection. Together these results provide an improved drowsiness detection model, highlight deficiencies in the current understanding of drowsy driving, and provide benchmarks for future predictive modeling analyses.