Learning lane change intentions through lane contentedness estimation from demonstrated driving

To assure a safe, comfortable and especially a cooperative driving experience while driving semi, highly or even fully automated, anticipation of the driving behavior of other traffic participants is needed. Because of the amount of different traffic situations and influence factors on the task of driving, due to uncertainties in environmental sensor measurements, and as a consequence of variable and individual driving styles, probabilistic models in combination with machine learning techniques are often applied to learn driving behavior from data. In this paper, with the help of a simulator study, the driving behavior of a subject group is examined regarding their intention to change lane on highways. The simulator is set up as a partially automated driving system that takes discrete maneuver wishes as input (lane change left or lane change right). If the traffic situation permits it, the requested maneuvers is executed automatically by the system. This generates ground truth labels that are being used to train a lane change intention classifier. The results show that the approach is able to predict upcoming lane changes at an average of more than 3.5 seconds in advance.