Image-based road network extraction approaches encounter a number of problems related to the lack of prior road data. In this paper, we present a road extraction approach using OpenStreetMap road vectors to accurately and robustly extract complete road information. Due to the persistent misalignment between image and map data, map registration is carried out first to move road vectors to corresponding image road centerlines. Junction templates derived from road vectors are matched with image-derived binary road mask and curvilinear response image to correct junction points. The non-junction point matching approach then effectively aligns the vector road network based on image curvilinear structures, within which width and orientation estimations are also embedded and can be obtained to recover the piecewise width of road segments. Road label mask is finally created with full road knowledge — centerline, width, connectivity, and topology. Our approach was tested on complex scenes and its effectiveness could be verified.