Understanding interactions between traffic participants based on learned behaviors

Florian Kuhnt, Jens Schulz, Thomas Schamm, J. Marius Zöllner

Abstract: Predicting vehicles' behaviors in a traffic scene can be very challenging due to many influences. Especially interactions with other traffic participants like vehicles or pedestrians are very crucial for the future movement while they are hard to model even with expert knowledge. In this paper we propose an object-oriented probabilistic approach that detects interactions between vehicles and is able to infer possible routes of traffic participants. Using the Object-Oriented Probabilistic Relational Modelling Language (OPRML), the interactions between vehicles can be modeled in an intuitive direct way. The probabilistic component allows Bayesian Inference on noisy sensor data and uncertain dependencies, while the object-orientation makes the model flexible to a varying number of traffic participants. Street-dependent as well as interaction-dependent motion models are learned from simulated situations and recordings of real traffic scenes. Finally, route prediction is evaluated at an exemplary intersection showing how the awareness of interactions reduces route prediction uncertainty and wrong predictions.