Segmentations of road area in high resolution images

This paper focuses on segmenting road areas in aerial images. The anticipated outcomes from segmentation are boundaries of roads and medians, and the ultimate goal of this research is to investigate how to build geo-specific road databases from aerial images for driving simulation. The digital line graph (DLG) from the United States Geographical Survey (USGS) is the starting point in locating roads. A method is presented to match DLG maps to high resolution aerial images that are not geo-referenced. The search space will be reduced significantly by using maps to eliminate unwanted areas. However the roads themselves produce images that are difficult to segment because of shadows, traffic, and discontinuity of pavements. By comparing histograms of hue gradient images, roads are segmented into two classes that are pavement and nonpavement (grass, sidewalk etc). Since segmentation can not give exactly the correct boundaries, edge detection is also employed in conjunction with the preliminary segmentation results. A rule-based system is developed to fuse these two types of data sets together to delineate a segmented road image