Estimating and predicting traffic situations over time is an essential capability for sophisticated driver assistance systems and 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. Building consistent probabilistic models of drivers interactions with the environment, the road network and other traffic participants poses a complex problem. In this paper, we model the decision making process of drivers by building a hierarchical Dynamic Bayesian Model that describes physical relationships as well as the driver's behaviors and plans. This way, the uncertainties in the process on all abstraction levels can be handled in a mathematically consistent way. As drivers behaviors are difficult to model, we present an approach for learning continuous, non-linear, context-dependent models for the behavior of traffic participants. We propose an Expectation Maximization (EM) approach for learning the models integrated in the DBN from unlabeled observations. Experiments show a significant improvement in estimation and prediction accuracy over standard models which only consider vehicle dynamics. Finally, a novel approach to tactical decision making for autonomous driving is outlined. It is based on a continuous Partially Observable Markov Decision Process (POMDP) that uses the presented model for prediction.