What contributes to driving behavior prediction at unsignalized intersections?

Safely passing through unsignalized intersections (USI) in urban area is challenging for autonomous vehicles due to high uncertainties of surrounding engaged human-driven vehicles. In order to achieve this, various variables have been selected to estimate and predict the surrounding human driver's behavior. However, it is still not fully clear what variables mostly contribute to driving behavior prediction at USI. This paper investigates the contribution levels of 13 features of human driver's decision-making at USI using a random forest approach. Thirty skilled driver participants are tested in a real-time driving simulator where the traffic scenarios with merging vehicles were designed in different motion styles to mimic real traffic. The experiment results indicate that the relative distance and velocity between merging vehicles have a wider contribution range (i.e., −0.4 −0.4) than the absolute velocity and distance features (i.e., −0.2 −0.2) to predict driver behavior. The contribution also varies over the selected feature values and driving conditions. This contribution research gains insight in the influence of different variables on driver behavior prediction at USI, thereby assisting researchers in selecting representative features in self-driving applications.