Autonomous driving has become the focus of autonomous navigation in recent years. Automotive industries seek the benefits of this research area for their purposes. Automotive navigation includes many aspects, one of them is the localization ability. Any information including stop signs, lane lines and road intersections can aid this localization. C urrent research attempts to develop a robust and efficient vision based roads and junctions detection on road networks using a monocular color camera mounted on top of a vehicle. A single color camera while gives enough information about the environment, is economical as well. Previous methods mostly use the combination of sensors which cost a lot. A few methods use monocular camera, and these methods suffer from assumptions which reduces generalization capability of the detection systems. The current work uses a combination of color, texture, coordinate and perspective features of road images, which leads to fewer assumptions and subsequently more generality, efficiency and robustness. The use of clustering method based on colors and coordinate features along with texture features have improved the robustness of our work. Our works come with two approach using clustering methods. In the first method, we assigned one cluster to road pixels. In the second one, we merge and select a subset of clusters to extract road pixels. The first approach needs homogenity assumption. To avoid this assumption, our main work developed with the second approach which results in more generality. The results of experiments demonstrate the robustness of the last proposed method in conditions such as shaded, ill-structured and non-paved roads. Keywords: Machine vision , Urban driving , Autonomous navigation, Road detection
