How to predict real road state from vehicle embedded camera?

Advanced Driver Assistance Systems (ADAS) based on video camera are increasingly invasive in todays car. However, if most of these systems work properly under clear weather, their performances drastically fall in case of adverse weather or bad lighting conditions. In this paper we study how to predict the road state: wet or dry. We simulate realistic embedded images relying on scene's physical data: road's Bidirectional Reflectance Distribution Function (BRDF), vehicle direction, sun position and daylight model of the sky. These data are used to produce a database of 640 synthetic images of wet and dry road scene, under different conditions (weather, date, direction). This database allows us to evaluate the relationship between conditions and road state, in order to determine from a given condition if the road state could be predictable. Finally, an optimization method is used to estimate road surfaces' BRDF parameters.