Robust approach for suburban road segmentation in high-resolution aerial images

摘要

The goal of this research is to develop an algorithm that accurately segments high-resolution images, where linear features, such as roads, are corrupted by noise. In high-resolution images, there are two types of noises that are obstacles to road segmentation. Noise could be within road areas (such as cars, water, differences in composite road surface mix) or unwanted contents outside road areas, like buildings and trees. To remove unwanted contents, the geographical information from the United States Geographical Survey (USGS) is used. The USGS provides a collection of road centre line information that has been collected for many years and can be used to limit the area for road segmentations close to roads. In this paper, a standard process was developed to align the USGS geographical information with the high-resolution images. USGS geographical data is used to eliminate background clutter that is disjunct from roads. The road segmentation process is then reduced to dealing with automobile traffic, shadows and pavement colour discontinuity within road areas. In order to achieve reliable road segmentation in the presence of these objects, the mean-shift clustering approach is used within the hue-saturation-intensity (HSI) space. Conditional morphological image processing techniques are also used to significantly improve the segmentation results. The proposed method results in the average accuracy of road segmentation above 85%.