Backward Segmentation and Region Fitting for Geometrical Visibility Range Estimation

We present a new application of computer vision: continuous measurement of the geometrical visibility range on inter-urban roads, solely based on a monocular image acquisition system. To tackle this problem, we propose first a road segmentation scheme based on a Parzenwindowing of a color feature space with an original update that allows us to cope with heterogeneously paved-roads, shadows and reflections, observed under various and changing lighting conditions. Second, we address the under-constrained problem of retrieving the depth information along the road based on the flat word assumption. This is performed by a new region-fitting iterative least squares algorithm, derived from half-quadratic theory, able to cope with vanishing-point estimation, and allowing us to estimate the geometrical visibility range.