The authors of this paper tackle the problem of distinguishing driver intentions in visual distractions to accomplish comfortable human-machine interactions. The discrimination targets are threefold: no visual distractions, i.e., looking directly ahead, and two types of visual distractions thathave opposite affects on safety, i.e., checking side blind spots and gazing at non-driving-related objects. The stochastic relationship between driver states and the three types of observations, or the driver's physical actions, artifact operations, and driving situations, is modeled with a Dynamic Bayesian Network. The experiments with a realistic driving simulator demonstrated how effectively of the proposed method enabled these driver states to be recognized.
Abstract