A Robust Gradient-based Building Area and Plane Extraction Method

Abstract

Extracting buildings from remote sensing data is an essential task in many applications. Currently, automatic building extraction from complex scenes is still challenging. This is due to the presence of dense vegetation in a site, variability in building sizes, materials and structures of building roof. For decades, researchers have continued to work on building extraction using two different kinds of data, i.e photogrammetric images and Light Detection And Ranging (LiDAR) data. The existing work has generally ignored the analysis of building material properties and has limited their study to the extraction of only those buildings with non-transparent roof. Thus existing methods are not effective in extracting buildings with a transparent roof. With their predefined size parameter settings, these methods might also fail in extracting small buildings. Furthermore, they overlook the importance of a local feature analysis in refining the building boundaries and eliminating the vegetation from complex scenes. For extracting the building details, i.e. the fine structures of roof planes, very limited attention has been paid to fusing/integrating the extracted features from the photogrammetric images and the LiDAR data. The novel methods we have proposed in this thesis aim to address these limitations. First, we have proposed a gradient-based building area extraction method which analyses the building properties, i.e. orientations and materials of buildings, from the photogrammetric images and the LiDAR data. This enables the proposed method to effectively extract buildings with a transparent roof. The proposed method also does not depend on the shape or size parameters as well; thereby it is more effective in extracting buildings of a larger size range. Next, we have proposed a building refinement process that utilises the local building feature analyses. It avoids many empirically set parameters such as height and shape thresholds. This enables the proposed refinement process to eliminate vegetation more effectively and to extract building portions at a lower height. Finally, we take advantage of the gradient concept in our proposed gradient-based building extraction method to extract the boundaries of building planes which are refined by using a new plane boundary regularisation process based on integration of the features from both input data. As a result, better refined structures (i.e. a smooth boundary) of building roof planes are extracted. In addition, the proposed method can also extract small roof planes as size or shape thresholds are not used. The performance of the proposed method is qualitatively and quantitatively evaluated on two different benchmark data sets. Our experimental results have shown that our proposed methods are effective in addressing the limitations described above and have outperformed existing state-of-the-art baseline methods.