Assessing map-based maneuver hypotheses using probabilistic methods and evidence theory

The prediction of the behavior of other traffic participants and the generation of appropriate motion hypotheses is a key capability of advanced driver assistance systems and autonomous vehicles. Motion prediction is a difficult task since it has to deal with the uncertainty within the environmental perception and the ambiguity of a traffic scene. For this reason we propose a two-layer situation analysis concept. This includes an associative and predictive situation model which combines probabilistic object hypotheses with a stochastic model of the road network in a curve coordinate system. Utilizing this description, we formulate various hypotheses regarding the evolvement of the situation using an Extended Kalman Filter supported by the Intelligent Driver Model. Furthermore, we introduce an evidence theory based situation interpretation to assess the several behavior hypotheses as well as to determine the inherent uncertainty. Especially in ambiguous situations, the ability to determine the imprecision by the difference of belief and plausibility of a certain hypothesis provides suitable information for an appropriate reaction. Both layers of the proposed situation analysis are not relying on training data and so it is not limited to previous known traffic scenarios. Finally, the capability of the concept is demonstrated by evaluating 157 maneuvers, recorded at an urban intersection.