Using Context Information and Probabilistic Classification for Making Extended Long-Term Trajectory Predictions

Intersections are among the most accident prone spots in traffic. Future Advanced Driver Assistance Systems (ADAS) are aiming to assist the driving task in these complex scenarios. This can be realized by assessing the criticality of possible occurring situations. For such criticality assessment techniques predicting the trajectories of the involved traffic participants several seconds in advance is necessary. In this paper we outline a method that makes exhaustive use of context information to reliably predict maneuver-specific trajectories up to 5 seconds into the future. Since the evolution of traffic scenes cannot be predicted with absolute certainty, approximating future states in form of probability density functions will be of great benefit in terms of robustness and reliability. We present an approach that is able to efficiently construct a discrete probability distribution by reformulating the problem as a probabilistic multiclass classification problem. The presented approach is evaluated on a real-world data set containing approaches to 85 different intersections. We show that we can make reliable maneuver-specific state estimations, even for a prediction horizon of up to 5 seconds.