Probabilistic Maneuver Prediction in Traffic Scenarios

Traffic scene understanding is an important part of future advanced driver assistance system (ADAS). In this paper we present a probabilistic approach for this problem. We use Hidden Markov Models (HMMs) for modeling the spatiotemporal dependencies of traffic situations. The parameters of the individual models are learned from a data base, generated by a professional driving simulation software. New oncoming situations are firstly perceived by an on-board sensor system including radar and camera. After the parameters of traffic participants are estimated they are used for recognition and prediction of the traffic maneuver by evaluating the likelihood for the learned models. Experimental results in real-world environments have shown that the method can robustly recognize driving maneuvers. As an example, the prediction of lane change maneuvers shows the benefit for different ADAS functions.