基於影像資訊與幾何原型重建三維模型之研究

Abstract

隨著影像建模技術的發展，人們可以藉由上傳對著物體所拍攝的不同角度的照片來獲得屬於自己的三維模型。但是這種基於影像資訊的方法容易會有相機校正的問題，導致最後產生的三維模型會有雜訊和部分扭曲的情形。我們的方法輸入一個點雲和多張照片，利用照片裡的線段資訊來解決點雲三維資料之雜訊和扭曲的問題。

在這篇論文中，我們提出的演算法分成三個階段。在第一個階段裡，我們從照片中擷取出的線段資訊來組成矩形平面。在接下來的兩個階段，我們從許多矩形平面中找出最能包覆點雲表面積的主要平面，然後利用區域和全域的對稱中心，產生更多的輔助平面。最後利用隨機抽樣一致的演算法包覆剩下的點雲表面。而使用者可以利用我們設計的介面對三維模型做調整。

我們提出一個方法可以處理影像品質不一以及在點雲邊界上的扭曲問題。我們所重建出的三維模型和單純只用隨機抽樣一致演算法所得到的三維模型相比，在物體邊界上有更準確的結果。

With the development of photogrammetric technologies, people can obtain their own three-dimensional models by capturing a set of photographs around the object. However, the image-based methods have the problem of imperfect camera calibration, and the result point cloud may have noise or distortion.

In this thesis, we propose a three-stage algorithm to reconstruct three-dimensional models from a point cloud and photographs. We extract the line segment information from input images to form rectangular planes, and apply an objective-function-based selection to find the primary planes. Due to the unexpected quality of input image, some regions may lack of line segments. Therefore, we further use the local and global symmetric properties to yield more auxiliary planes. Then, we apply a RANSAC method to fit the rest point cloud. We also allow users interactively refining the reconstructed model.

We propose a novel method to deal with the distortion at the edge of input point cloud and the unexpected quality of photographs. Our result models are more precise than the results which directly apply the RANSAC method on the point cloud.