Automatic Road Extraction Based on Multi-scale Modeling, Context, and Snakes

This paper approaches the problem of road extraction from three different directions. The first is the use of multiple scales. This combines detailed information of fine scale, like the markings, with abstract information of coarse scale, like the road network. The second direction is the extension of the multi-scale modeling with the context, i.e., the relations to other objects like buildings or trees. The context itself is split hierarchically into local context sketches, like occlusion shadow, which is modeling a tree casting a shadow on the road, and global context regions, i.e., open rural, suburb urban, and forest areas which comprise the whole image. The context information is very useful to focus the extraction. The third direction taken in this paper is the use of snakes. So-called ribbon snakes are used not only to extract roads in a robust manner in fine scale, but they can be also used to bridge gaps in the extracted roads due to occlusions or shadows cast by buildings and trees. Practical examples show the validity of the approach.