Forecasting Markers of Habitual Driving Behaviors Associated With Crash Risk

Both distracted and aggressive driving are habitual in nature, constituting an insurance risk, which has been difficult to quantify. Here, in this paper, we propose a method that produces short term predictions for these two dangerous driving behaviors. The method feeds an Extreme Gradient Boosting (XGB) algorithm with the most informative features of a set of physiological and vehicular variables. The XGB algorithm operates on a learning window covering the last 30 seconds to make fast track predictions (FT) for the next 10 seconds. For aggressive driving, FT predictions are final, while for distracted driving, FT predictions are weighted over one minute, to form a meta-prediction. This more deliberative process for predicting distractions fits their intermittent manifestation. The method has been tested on SIM 1, a publicly available dataset from a distracted driving experiment. In this dataset, the drivers ( $n=59$ ) are labeled as distracted based on the presence of mental activity or physical interactions antagonistic to the driving task; their driving style is defined by steering and acceleration, and is classified as aggressive or normal. The method attains classification performance that exceeds 87%. Alerting drivers when distractions and aggressiveness have taken hold on them can provide sobering awareness, given that people drift into these states subconsciously. The behavioral modification effects of such awareness mechanisms are rooted in Cognitive Behavioral Theory. The proposed method can also be used in future vehicles with advanced automation, weighing in the computer's decision to wrest vehicular control from an unrepentant driver.