近年由於三維房屋模型有大量且廣泛的應用需求，利用光達點雲重建三維房屋模型是其中一項備受關注的技術，並有許多研究單位投入人力與資源於此項技術的演算法開發與流程優化。在諸多尚未解決的議題中，如何提升點雲資料的品質使其重建模型更具完整性即為本研究欲解決的重點之一。例如，處理在儀器掃描的過程中，由於房屋受到自身結構或外在物體的部分遮蔽問題。而使用不完整的光達資料重建三維建物是另一項具挑戰性且實用性高的課題，因此，本研究提出重建表面的空洞填補 (hole-filling) 機制的點雲資料處理演算法，能在資料缺失的部分獲取正確且具有代表性的點雲。另外，本研究所研發自原始點雲中萃取邊緣點與建物邊界線的演算法，亦能有助於利用地面光達點雲重建精確幾何的三維房屋模型。 主要研究內容包含以下步驟，首先以人工移除非房屋的點雲，並將其分割成數個區塊。接著，利用各區塊的點雲資料判斷是否有屬於房屋表面的門或窗等形態的表面間隙 (surface gap) 。偵測到的區塊利用邊界線與表面間隙可再切分成子區塊，其中的邊界線擬合自點雲中萃取的邊緣點。在排除掉表面間隙後，剩餘的間隙則可視為資料缺失所造成的空洞。 為了填補點雲上的空洞，將各區塊的點雲投影至該區塊坐落的主平面上，並將其轉換成網格格式。接著利用最近相鄰內插法填補外包多邊形 (convex hull) 中空白的網格資料，其中網格內若有複數點則以其各點的幾何中心做為代表。最後可建置成多邊形網格表面模型，再以人工方式調整與優化。實驗案例成果顯示，本研究提出之演算法能處理平面或曲面的房屋，且對於地面光達點雲能有更有效率與準確地重建出三維房屋模型。 ;Reconstructing a 3D building model from LiDAR point cloud has recently gained attention. This is probably due to a large array of potential applications. Several Researches have been conducted on this topic and lots of algorithms have been accordingly developed. Unfortunately, despite the extensive research works, there are still many problems unsolved. One of the main challenging issues yet not thoroughly studied is how to improve the quality of the data in terms of completeness. For example, during the scanning process, parts of the building may not be sampled, either because they are occluded by other objects or they are self-occluded. Reconstructing a building from such incomplete data is not a trivial task. This research presents an approach for accurate reconstruction of geometric 3D building models from terrestrial LiDAR point clouds. As a main focus of the research, an algorithm is proposed for obtaining correct representative data points for the missing parts of an incomplete data set – a task known as hole-filling in surface reconstruction parlance. Further, an effective method for extracting building edge and boundary points from raw point cloud data is developed. The proposed reconstruction method is in multiple stages. After preprocessing the data to remove unwanted points and outliers, point cloud is segmented into regions. Each resulting segment is checked for containment of surface gaps such as doors and windows. When such a segment is found, it is further decomposed into simpler sub segments using the lines corresponding to the boundaries of the included surface gap(s). The line segments used in decomposition process are obtained from the fitting of lines to the extracted edge points. This systematic exclusion of surface gaps from segments allows all remaining gaps in the data to be treated as holes. To fill holes, each segment is projected onto its underlying principal plane and converted into regular grid format. The empty grids that are within the convex hull of the projected data are filled using nearest neighbor interpolation. Grids with more than one point are represented by the centroids of the occupants. After these processes, a polygonal mesh surface model is generated which is further adjusted manually and then refined. The results of the test case demonstrate the capability of the proposed approach for effective, accurate and high quality reconstruction of 3D building model from terrestrial LiDAR point clouds. The proposed reconstruction method can handle building with both planar and curved surfaces.
