本論文研究包括空間點雲重建與自動建模。在空間點雲重建部分，為解決點雲因掃描場景紅外雜訊干擾、反射、透明表面造成資訊遺失的問題，我們提出遞迴低通延伸點雲技術來重建點雲完整樣貌；在抗雜訊分析部分，遞迴低通延伸技術具備低通擴張特性，對雜訊有一定的容忍度，訊雜比低的點雲仍有良好的準確度與精確度。 本論文亦提出自動建模技術，結合二維影像邊界萃取、深度學習語意分割與模型假設，我們成功地從空間點雲萃取物件資訊並自動建模，輸出模型交換格式DXF (Drawing Interchange Format)。;The thesis presents a study containing topics of point cloud complement and auto-modeling. In order to solve the problems including strong noise from shiny, infrared source, reflecting or transparent surface, and strong absorb materials, which cause information loss and the defect of the point cloud and, we proposed to use iterative low-pass pervasion method to complete depth images. The experiment result shows that with strong noise interference, iterative low-pass pervasion method still has good accuracy and precision. We also study auto-modeling technology. With boundary extraction from RGB Images, 2D image semantic segmentation, and hypothesis of model, we successfully extract model information from point cloud and then transfer it to DXF (Drawing Interchange Format) .
