Integration of multi-source data for automated building extraction

Abstract

Digital building models are used in various applications and are traditionally acquired by manual digitization of aerial images in stereo view using photogrammetric stereoplotters. Manual digitization is a tedious and time-consuming task, which requires skilled operators and expensive equipments. Therefore, methods that can assist the operator in performing the entire or part of the task automatically are of great significance. Automated methods for extraction of buildings face a number of problems. Semi-automated approaches involve a considerable amount of interaction. Fully automated approaches that work solely based on a single source of data, suffer from the lack of robustness due to complexities involved in data as well as in buildings. In this thesis, a new approach is presented for semi-automated extraction of buildings from a single aerial image with minimum interaction. In this approach, a fuzzy grouping method is developed to select the most significant group of image lines. For the integration of data from multiple sources, two new approaches are presented in this thesis. These approaches are based on a split-and-merge process, which is developed in this work for the refinement of image segmentation by integrating image and height data. The fuzzy grouping developed in this thesis is based on a number of fuzzy rules that define the significance of proximity and parallelism relations between image lines. The significance measures are computed within a fuzzy inference system, and stored in a weighted graph structure. A search method is used to find a group of most significant image lines in the graph. Experimental evaluation of the fuzzy grouping method shows that this method results in better groupings as compared to the existing grouping methods. In the semi-automated approach presented in this thesis, a model-driven strategy is devised, which is based on the fuzzy grouping method and a modified model matching technique. In this strategy, the model is interactively selected by an operator and is fitted to the image automatically. Image lines are extracted and grouped using the fuzzy inference system. The correspondence between the image and model lines is established using an intelligent tree-based search implemented in Prolog. Corresponding lines undergo a matching procedure, which determines whether or not a match can be found between the given model and the image lines. The split-and-merge process developed in this work is a robust method for refining image segmentation by making use of height data. In this method, the image is segmented, and a robust plane fitting process is applied to fit planar surfaces to height points belonging to each image region. The detected planes are used to split overgrown regions and merge undergrown regions. By the integration of image and height data in this method, parametric forms of roof planes can be reconstructed. Experimental evaluation of the split-and-merge process indicates the successful performance of the method in refining the segmentation results. Based on the split-and-merge process, a fully automated approach to building extraction from image and height data is presented in this thesis. In this approach, the number and attributes of the detected roof planes are used as indices to a library of parametric building models, to generate a number of roof hypotheses. These hypotheses are verified using the model matching method, and the parameters of the accepted model are computed. The split-and-merge process is also used in another new approach to fully automated building extraction from image, height, and 2D map data. In this approach, buildings are detected using the ground plans and height data. The split-and-merge process is applied to fuse image and height data, and derive the parametric forms of the roof planes. Walls are reconstructed as vertical planes upon the ground plans. A plane patch reconstruction method is developed to intersect model faces, and assemble the resulting plane patches to form a generic polyhedral model. Results of the experimental evaluation of the proposed approaches indicate the promising performance of the approaches in terms of such criteria as success rate, accuracy and computational efficiency. The semi-automated approach was successfully used for the reconstruction of 4 buildings with an accuracy of better than 1m. The two fully automated approaches were applied to 14 buildings. The first approach achieved a success rate of 71%, while the second approach achieved a success rate of 78%. The accuracy of the reconstructed roofs was found to vary between 6cm and 48cm in both approaches. The computation time was about 5 seconds per building in all the proposed approaches.