Integrated modeling of driver gaze and vehicle operation behavior to estimate risk level during lane changes

In this paper, we investigate a method for detecting risky lane changes using integrated modeling of driver gaze and vehicle operation behavior. Driver gaze direction and vehicle operation behavior are broken down into discrete acts, e.g., looking in the rear view mirror, braking, etc., and sequences of these actions are jointly modeled using multi-stream hidden Markov models (HMMs). Driving data is recorded on expressways as drivers pass leading vehicles, i.e., the drivers make two lane changes, first to pass leading vehicles and then to move back into their original lanes. Since actual driving risk levels are difficult to measure, the risk level of each lane change is rated by subjects, and we assume their scores represent the “ground-truth” risk level. By jointly modeling gaze and vehicle operation behavior, we improve the performance of risky lane change detection. To more accurately evaluate overall risk, we use data from multiple lane change events. We obtain an average correlation coefficient of 0.80 between HMM likelihood scores and subjective risk evaluation scores by accumulating HMM likelihoods for a period of fourteen minutes. By accumulating these indicators for the previous twenty minutes, a 96.0% risky driving detection rate is achieved with a 7.1% false positive rate.