模型化的程序中，應該考慮建物邊緣的處理。影像中特徵線提供了良好的線索以找尋房屋邊緣。因此透過三維點雲與影像上的特徵線分析以判定房屋之輪廓。本研究萃取房屋輪廓，並以房屋輪廓當作約制精化數值地表模型。實驗成果顯示，結合半全域匹配法與中左右視窗匹配法可以達到互補之效果並提升匹配品質，另外以線特徵為約制精化模型的方法可以增進最後模型成果的品質。 ;Digital Surface Model (DSM), which describes the surface topography, is an important data source in geoinformatic applications. Considering its importance, this research uses aerial images to construct DSM for building areas. This study includes two major works：(1) point cloud generation and (2) surface modeling for DSM reconstruction.
It is a practical way to generate 3D point clouds by matching multiple images. Because the density of 3D points clouds may influence the constructed details of DSM, this research employs dense matching method to generate 3D points clouds. Matching methods using local single target window only consider local similarity near the matching points. It lacks global links to other pixels. The matching results could be improved, provided that the global similarity is considered. Semi-Global Matching (SGM) considers connected paths with smoothness constraints and combines local and global image information so it can get stable results. However, smoothness constraint might be unable to cope with the matching ambiguity in the area with surface discontinuity. Central-Left-Right Matching (CLRM), on the other hand, considers local feature constraint using multi-windows to increase matching quality around feature regions. Thus, the integration of CLR and SGM is proposed in this investigation. Object-based image matching starts from a groundel to connect related image pixels. Because object-based image matching strategy can connect multiple images with different image resolutions, it will be employed in this research.
In the DSM modeling, surface discontinuity should be taken into account. Feature lines in the images provide a valuable clue for the detection of possible surface discontinuity such as at building boundaries. Thus, 3D break lines might be determined by incorporating the point clouds and image feature lines. In this research, we extract building boundaries followed by the inclusion of those boundary lines as constraints to shape the DSM. The experimental results indicate that the integration of CLR and SGM can increase the quality of image matching. In addition, the proposed method that uses feature constraint to shape DSM can improve the quality of the generated DSM.
