Research On Building Reconstruction Based On High-resolution Aerial Stereo Images

The 3D models of buildings have a wide range of application, such as city planning, building survey, communication construction, virtual travelling, heritage protection, disaster monitoring and reconnaissance, precision strikes in military. Building reconstruction based on high-resolution aerial stereo images is one of major ways to generate the 3D models of buildings. Many researchers have studied on automatically reconstructing buildings, and much large progress has been achieved, however, some key problems remain unsolved. Therefore manual interventions are still essential for many existing methods. After analyzing existing advanced solutions, the paper studies on the problem of generating automatically the accurate 3D models of the buildings with complex polyhedron-shaped structure by using high-resolution aerial stereo images, and achieves the following innovations.(1)A suit of scheme for generating the 3D models of the buildings based on high-resolution aerial stereo images is devised and implemented. This scheme is comprised of five steps such as extracting lines, line stereo matching, generating 3D models, correcting 3D models and texture mapping, and can be used to reconstruct the buildings with the roofs of the simple and complex gable and multiple spires.(2)A method which can filter the false line features based on image content is proposed. Based on the application background of building reconstructing, in the existing line extraction algorithm two strategies are added up as the posterior steps of edge detection and line extraction. According to the different characteristics between the true and false lines in the image, this method can enhance the ability of extracting poor line features and preserving true ones, but also choke back the false line features effectively.(3)An algorithm of line stereo matching for solving the problems of partial occlusion and feature uncertainty is proposed. Partial occlusion and feature uncertainty are two critical problems that make the problem of line stereo matching more complex. Partial occlusion is an explicit problem involved in the process of projecting spatial line segments, however, feature uncertainty is a random one involved in the process of extracting the projection of spatial line segments. Although both of them can bring out the complex correspondences among line features extracted from various images, such as one-to-many and many-to-many ones, their influence on stereo matching is different and should therefore be solved by different methods. However the existing methods did not distinguish them. In this paper the different solutions for partial occlusion and feature uncertainty are given. By combining the two solutions a new framework of line stereo matching is proposed and furthermore some critical questions involved in it are answered. Employing the projection theory, a new concept of feature group is developed, which is comprised of all the projective lines of a spatial line or a special line structure in both the left and the right images, and then a projective line match relationship graph, which describes the relationships of correspondence, compatibility and combination between the projective lines, is developed to model the occlusion problem. Employing the statistical theory, the uncertain mapping relationship between the lines extracted from images and their homonymous projective lines is analyzed, and then a statistical model for restoring the homonymous projective lines with the extracted lines is built. Based on the above two models, a new method for line stereo matching is developed. Firstly, all extracted lines are used to restore their potential homonymous projective ones using the statistical model of line feature uncertainty. Secondly, a graph is built, each one of whose nodes is a potential match pair of restored projective lines from the different images, and whose edges are linked according to the compatibility relationships and combination relationships between the restored projective lines contained in the every two nodes. Thirdly, all maximal cliques of this graph are searched and then the energy which describes the probability of a maximal clique corresponding to a real spatial line or spatial line structure is measured. Lastly, all appropriate maximal cliques are selected by an optimization algorithm according to the compatibility among all the maximal cliques and their energies. The experimental results applied on the real stereo images prove the validity of the method.(4)A new method is proposed to automatically generate the surface models of the buildings shaped by polyhedrons. Planar patches and the topological relationships between them are the major information of the surface model of a polyhedral building. This method proposed here can construct the completed borders for all the patches and build the topological relationships simultaneously. It is implemented in three phases as follows. In the first phase, an algorithm for detecting and positioning the patches in the roofs of buildings is developed based on the strategy of hypothesis-verification. The algorithm includes four steps. Firstly, based on the results of matching lines, the 3D lines are reconstructed and then the ones with higher confidence level are chosen to generate all the plane hypotheses. Secondly, every plane hypothesis is segmented with the 3D lines that are judged being in this plane and then some sub-patch hypotheses are obtained. Thirdly, some credible sub-patches are chosen according to the photometric similarity of the sub-patch hypotheses across the two images and the compatibility relationships between them resulted from occlusion. Lastly, the open planes used to construct the surface model of buildings'roofs are generated and their scopes are delineated approximately. In the second phase, a method for constructing the borders of the open planes and the topological relationships between them is proposed, which includes three steps. Firstly, the intersecting lines are computed with the above open planes, based on which the possible topological relationships are assumed and the lost planes are detected and complemented according to the information from images. Secondly, all the borders of the open planes are constructed by the intersecting lines, the 3D lines, unassigned 2D lines and some simple heuristic formulas according to their confidence levels in the descending order. Lastly, the wholly optimal surface model of the roof is computed by fusion of two rules such as the completeness of the borders and the compatibility of the topological relationships. In the third phase, the complete surface model of the whole building is generated by filling the upright walls.(5)An algorithm for correcting the surface models of the polyhedral buildings is proposed. It is inevitable that there is error in the generated 3D models because of the errors in extracting lines and inaccurate camera parameters. This algorithm will be used to improve the accuracy of the surface models. Firstly, all the regular geometrical relationships, such as parallel, perpendicular and asymmetrical ones, among the 3D lines in the contour of every patch are detected. Secondly, a cost function whose variables are the pose parameters of the surface model is defined based on the specified geometrical constraints, and then the model-shift method is employed to compute its optimal pose parameters.