Quantifying visual road environment to establish a speeding prediction model: An examination using naturalistic...

Speeding is one of the major contributors to traffic crashes. To solve this problem, speeding prediction is recognized as a critical step in a pre-warning system. While previous studies have shown that speeding is affected by road environmental design, research in predicting speeding behavior through road environment features has not yet been conducted. Furthermore, there is a large discrepancy between actual and perceived road environmental information given that a driver's visual perception plays a crucial role as the dominant source of information in determining driver's behavior. Thus, this paper aims to establish a speeding prediction model based on quantifying the visual road environment to improve the design of pre-waring systems, which can predict whether drivers are going to speed and provide them with visual or/and audio warnings about their current driving speed and the speed limit prior to the occurrence of speeding behavior. Twenty input variables derived from three categories including visual road environment parameters, vehicle kinematic features, and driver characteristics were considered in the proposed speeding prediction model. Especially, the road environmental design factors consisting of the visual road geometry and visual roadside environment as perceived by the driver's eyes were quantified using a visual road environment model. Field experiments were conducted to collect naturalistic driving data concerning speeding behavior on the typical two-lane mountainous rural highways in five provinces of China. Random Forests, an ensemble learning method for regression and classification, were applied to build the speeding prediction model and variable importance was calculated. Additionally, logistic regression was used as a supplement to further investigate factors impacting on speeding behavior. A speeding criterion was defined with two levels in this study: a lower level (exceeding the posted speed limit) and a higher level (10% above the posted speed limit). Under both levels of the speeding criterion, the speeding prediction model performed well with high accuracy (over 85%). This model could use the value of the variables obtained from the current position to predict drivers' speeding behavior at the future position located a sighting distance away. This interval was sufficient for a pre-warning system to give a speeding warning that a driver with normal perception-reaction time (around 2.5 s) could respond to. Findings in this study can be used to effectively predict speeding in advance and help to reduce speeding-related traffic accidents.