Mixed road surface model for driving assistance systems

The problem of road/obstacle separation of 3D points is revisited in this paper, in the context of driving assistance systems and dense stereovision. We propose some measures to cope with scenarios where only few road points are recovered in 3D, or the road has an atypical geometry. When a non-planar road model is employed, e.g. a quadratic road surface, the vertical profile of the road is usually reliable for the 3D region that contains the road inliers. Therefore, using this surface for obstacle/road separation outside the inliers region is less reliable. Our solution is a mixed road model: the quadratic road surface is filtered temporally and extended with a planar patch to cover the whole space of interest. The planar surface used for extension is computed from the parameters of the quadratic surface.