利用環景圖建構二維平面圖

Abstract

我們發展一個利用全景圖半自動建構二維平面圖的方法，其目標是建立室內公共空間的平面圖。現今的數位戶外地圖已經非常普遍，但當人們進入新的建築物時也需要室內地圖，有名的建模方法像是運動恢復結構法通常需要大量和強健的特徵點來建立，但室內環境卻比較缺乏特徵點使得無法有較佳的結果，因此我們延伸建構平面圖的演算法[1]來克服這個問題。先讓使用者在全景圖上標示轉角的位置，再利用轉角與全景圖中心所形成的射線相連轉換成平面圖形，接著利用同樣方法依序產生多個平面圖形連接合成完整的平面圖；然而上述方法將無法自動的找尋相對應的牆面和被限制於曼哈頓世界假設，使無法結合成完整精確的平面圖，為了要解決上述問題，所以提出了兩大方法：牆面對應法和平面圖修整法。牆面對應法會在每兩張全景圖中所標示出的牆面找出對應，以利接下來平面圖形相連接成平面圖，我們利用幾何與圖像上的相似性來完成；在平面圖形結合完之後會因為生成與接合時留下許多錯誤，因此需要利用平面圖修整的技術來對整體的平面圖做完整的修整，我們將利用最小平方法將點和線之間距離最小化來幫助整體的修正。透過上述兩個方式，我們可以產生室內平面圖並且可建立非曼哈頓與少量特徵點的場景，最後我們利用幾組具有挑戰性的測試資料並得到與實際的平面圖相似的結果。

This thesis proposes a semi-automatic method that generates a 2D floorplan from cylindrical panoramic images in indoor public area. Nowadays, digital maps are common for outdoor, but people also need indoor map when entering a new building. Popular reconstruction methods like structure form motion (SfM) often need sufficient and robust features, which are lacked in many indoor environments. Our purposed method is improved by the shape generation algorithm [1] can overcome featureless conditions. The input panoramic images are separated by corner selection from a user and generate corner rays from the panorama centers to corner positions. The shape generation algorithm connects adjacent corner rays to generate shapes, and then the shapes are combined to the floorplan. Nevertheless, the said algorithm has some problems; for example, it cannot find the correspondence to combine and is limited to the Manhattan world assumption. In order to solve these problems, we present two methods: wall matching and floorplan refinement. Wall matching will correlate the wall correspondence that the wall is separated by corners. We apply the geometry and image similarity to find the best wall correspondence, and then combine the shapes to the floorplan. However, when our method computes shape combination, the error is generated from shape generation and propagates to next shape, leading overall floorplan to the inaccurate result. Therefore, we develop the corner-line minimization in least square method that is a robust floorplan refinement algorithm to get more accurate result. Finally, the method can be used to generate indoor floorplan in featureless and non-Manhattan world scene. We also demonstrate results on several challenging datasets, and the results of floorplan are similar to ground truth.