Toward an automatic extraction of the road network by local interpretation of the scene

This paper deals with the automatic detection of road network from aerial images. Our approach can be characterised by a two-level processing: firstly, a low-level knowledge is used to extract most of the roads, then semantic information is applied to solve interpretation problems. As we aim to implement this system in a production framework, we have been very careful about the reliability of the result. According to this objective, an incomplete but very reliable network is better than a complete one but that contains also many misdetections. The results are rather fair: we can assess that more than 60% of the road network could be automatically extracted. Nevertheless, some cases will still require a human interpretation or the help of complementary data because of the complexity of some problems.