Automatic Road Extraction through Light Propagation Simulation

We present a road extraction system from aerial imagery guided by an imprecise road network. After a general discussion on road extraction strategies, we focus on the road recognition module we propose. It relies on two principal elements : a general purpose parallel structures detector in images, which uses a simulation of light ray propagation and reflections; and a dynamic programming based algorithm which constructs a network of hypothetical roads from the previously detected parallelism. We then use a voting scheme involving numerous criteria to select the most probable run of the road. We show through a set of examples that this system works well in rural areas, and also has promising possibilities for the interpretation of complex situations, especially those of multiple lanes roads, like motorways.