Learning Context Sensitive Behavior Models from Observations for Predicting Traffic Situations

Estimating and predicting traffic situations over time is an essential capability for sophisticated driver assistance systems or autonomous driving. When longer prediction horizons are needed, e.g., in decision making or motion planning, the uncertainty induced by incomplete environment perception and stochastic situation development over time cannot be neglected without sacrificing robustness and safety. Especially describing the unknown behavior of other traffic participants poses a complex problem. Building consistent probabilistic models of their manifold and changing interactions with the environment, the road network and other traffic participants by hand is error-prone. Further, the results could hardly cover the complete diversity of human behaviors. This paper presents an approach for learning continuous, non-linear, context dependent process models for the behavior of traffic participants from unlabeled observations. The resulting models are naturally embedded into a Dynamic Bayesian Network (DBN) that enables the prediction and estimation of traffic situations based on noisy and incomplete measurements. Using a hybrid state representation it combines discrete and continuous quantities in a mathematically sound way. Experiments show a significant improvement in estimation and prediction accuracy by the learned context dependent models over standard models, which only consider vehicle dynamics.