Abstract: Quickly extracting road networks from high-resolution remote sensing images is crucial in mapping, urban planning, and GIS databases updating. Semi-automatic road extraction, as the main method of road surveying and mapping, is a labor-intensive task. In order to reduce the cost of manual intervention and improve extraction efficiency, this paper proposes a fast road centerline extraction algorithm based on geodesic distance field. First, the optimal circular template is proposed to automatically estimated the road width and adjust the manual seeds to road center based on the morphological gradient map, and the road saliency map is calculated according to the local color features inside the templates. Second, we propose the soft road center kernel density based on road saliency map which overcomes the difficulty of threshold presetting of road segmentation in traditional road center kernel density estimation. Most importantly, a geodesic distance field is proposed to quickly extract the geodesic curve between two consecutive seeds, which dramatically increase the efficiency of our algorithm. Finally, we introduce the mean filter into our scheme to smooth the road centerlines. Extensive experiments and quantitative comparisons show that the proposed algorithm can greatly reduce manual intervention without losing much accuracy, and significantly improve the efficiency of road extraction. Furthermore, the proposed algorithm takes almost the same time to extract any length of road centerline given fixed image size, and no hyperparameters need to be set. The algorithm behaves good experience in human-computer interaction.

Key words: geodesic distance field, curve smoothing, road centerline extraction, remote sensing images