Modeling Driving Behavior at Single-Lane Roundabouts

Kurzfassung

Roundabouts are considered important because converting an intersection into a roundabout has been proven to improve safety. However, the absolute number of crashes at roundabouts is still high. Many crashes occur because car drivers fail to yield. Intelligent systems can increase safety if they can prevent crashes by precisely predicting driver maneuvers. Therefore, a reliable and trustworthy predictive model of driver maneuvers is needed. A few studies analyze human behavior at roundabouts. However, they focus on an operational timescale rather than on maneuvers on a tactical timescale. Tactical maneuvers have mostly been investigated in scenarios about typical intersections and overtaking. Thus, there is still a lack of research on driver maneuver prediction at roundabouts. To fill this gap, the objective of this thesis is to develop a model that can predict driver maneuvers at single-lane roundabouts. Two types of driver maneuvers are possible in front of each exit of a roundabout: exiting the roundabout and staying in the roundabout. To predict which maneuver a driver will execute in front of an exit, a driver maneuver predictive model was developed on the basis of an analysis of driver behavior data acquired from a field study and a simulator study. Soft-classification algorithms were proposed to train the predictive model. The model consisted of four sub-models for four different scenarios, which were defined by the correlation between roundabout layouts and drivers’ steering behavior. The sub-models make it possible to predict the exiting or staying maneuvers executed in the corresponding scenarios. Furthermore, a personalized predictive model was developed to adapt to individual drivers because different drivers have different driving styles. The driver maneuver predictive model shows excellent predictability: In the scenarios without traffic, the model reported prediction results for more than 97.60% of test drives at the position 10 m from the exits of the roundabouts. Of these drives, more than 97.10% were predicted correctly. The personalized predictive model provided even better prediction results for individual drivers with significantly different driving styles. Moreover, the driver maneuver predictive model also successfully predicts drivers’ maneuvers in most scenarios with cyclists. The prediction results show that steering angle, steering angle speed, velocity, and acceleration of the ego car provide the most important information. With this information, the model can predict the maneuver of a driver with any type of driving style at a single-lane roundabout with any type of layout.