We introduce a lane marker detection algorithm that integrates 3D attributes as well as 3D relations between local edges and semi-global contours in a Bayesian framework. The algorithm is parameter free and does not make use of any heuristic assumptions. The reasoning is based on the complete conditional probabilities of the different cues which are estimated from a training set. The importance of the individual visual cues can be computed using a standard measure and the cues can then be combined in an optimal way. In addition we show that when doing 3D reasoning, the uncertainties connected to the reconstruction process need to be taken into account to make the reasoning process more stable. The results are shown on a publicly available data set.