| dc.description.abstract | Nowadays, three dimensional data has been widely used in the field of geospatial information. To reconstruct 3D GIS, major components include buildings, roads, utilities, vegetation, etc. Among those components, building models are the most prominent in the reconstruction work. Aerial images are commonly used in building reconstruction. In most of man-made scenes, line segment features usually exist along boundaries. Thus, line matching plays an important role in the building reconstruction. The building reconstruction with line matching algorithm is performed using high overlapping aerial images in this study.
The proposed method includes (1) boundary determination and (2) roof structure reconstruction. If the building boundaries can be determined, then the roof structures can thus be processed and the interference of non-target objects can also be reduced. The basic procedures in those two parts are similar with the differences of parameter settings. First, feature extraction is to extract the straight lines, and then use line matching with the strategy of left-right windows to locate the conjugate lines in multiple images. In the 3D line positioning, space line intersection with the blunder detection are combined to derive quality results. The reconstruction and refinement processes are done in an iterative way.
In the boundary determination, for the reason that area of interest includes non-target objects, the interest line selection is needed for the estimation of their features. In the roof structure reconstruction, the areas of interest are produced by the back-project of the building boundaries. All of the extracted lines are selected in this stage to avoid line missing. After the 3D line positioning of the roof structure, the boundary model and roof structure are integrated with the initial model reconstruction and refinement to generate the 3D building model.
The experimental results indicate that multi-angle images can improve the matching successful rate. With the blunder detection, unreliable matched line segment can be eliminated effectively. The RMSE of the test models can reach 0.1m in X and Y direction, and 0.5m in Z direction. | en_US |