Large-scale efficient extraction of 3D roof segments from aerial stereo imagery

More Info
expand_more

Abstract

3D reconstruction of the built environment is a widely studied Geomatics topic. Resulting city models are used for a great variety of purposes. The reconstruction of roofs remains challenging. These roofs are not only building blocks for the city models but are of direct interest as well. The company READAAR uses roof segments for the detection of asbestos and PV potential analysis. Typical clients for such analysis are muncipalities and provinces. Currently, READAAR extracts roof segments from the gridded LiDAR dataset: algemeen hoogtebestand nederland 2 (AHN2). The use of LiDAR data however comes with some limitations. Most importantly, outside the Netherlands, LiDAR data is not always available as it is relatively expensive to gather. Furthermore, the point density of the AHN is 6-10 points/m^2, which limits the amount of detail that can be extracted. In the Netherlands countrywide aerial stereo imagery is available at a resolution of 10cm, this potentially gives a point density of 100 points/m^2 after image matching. This research explores the possibilities of using aerial stereo imagery instead of LiDAR data for the efficient large-scale extraction of 3D roof segments. A workflow is designed in which stereo matching and extraction of segments are integrated. This makes the workflow both efficient and easily scalable. Roof segments are extracted in two steps. First, a watershed segmentation is applied to retrieved color segments. Second, these color segments are clustered based on their orientation. The resulting roof segments generally have a higher quality than the segments retrieved with READAARs LiDAR-based approach. However, problems do occur, especially in shaded areas. This could possibly be solved by integrating LiDAR data when available. Another recommendation for future research is improving the matching by using multi-view matching or possibly neural networks. Furthermore, the segmentation could potentially be improved by using multiple images from different years and processing building blocks.