A Study on Scene Registration and Patching Algorithms for Large-Scale Virtual Reality

虛擬實境應用於醫學、軍事、工業上等等的教育訓練，幫助操作者進行沒有時地限制的反覆練習，以便能完成在真實世界中困難或失敗代價高昂的工作。虛擬環境的建造是虛擬實境領域中最常有的應用之一，然而欲建構戶外場景的大型虛擬實境場景，除了以人工繪製以外，商用的套裝系統皆索價高昂。本研究提出了一個新的流程用以建構大尺度虛擬實境，主要的貢獻是以低成本、快速、的逆向工程來完成建構大尺度虛擬實境，結合環場影像以及雷射掃描空間資訊建構大尺度虛擬實境模型。可針對校園環境格局做實景建模，並以真實影像做為貼圖材質。建立模型的五步驟分別為：點雲及影像資訊擷取、單一場景建構、場景連結、修補、貼圖。首先，以雷射測距儀與數位單眼相機同時取得環場空間資訊以及紋理影像，配合取景限制簡化建構模型的程序。在連結場景的過程中，過去全域比對法僅能得到唯一解，易受到沒有良好初始值以及場景不均勻的複雜點雲資訊之影像，本研究提出創新的方法以鳥瞰方向的Z-buffer結構推估空間資訊車在場景與場景之間的移動關係，以獲得資訊車位置之機率分布，並且以空攝影像做為場景接合的全域限制以及修正基礎，在本研究中場景與場景之平均距離為24.2公尺，接合成功率提高至85%。本研究提出創新的貼圖方法，能克服傳統三維模型貼圖需人工給定材質與位置的問題，以取景位置對場景模型做分割，由取景中心投影全景影像至場景模型。Virtual reality has been applied in many fields to assist people in dealing with hard problems in the real-world. Virtual environment construction is the most common application in virtual reality fields. However, large-scale environments such as campus environments containing many buildings are usually reconstructed by manual operation or through very expensive instruments. Therefore, we developed a novel system for building large-scale virtual reality models with photorealistic appearance. There are five steps of model reconstruction: data collection, single scene reconstruction, scenes registration, scenes repair, and texture mapping. First of all, a laser range finder and two digital single lens reflex cameras are integrated to get space information and object texture images simultaneously. Besides, the environment model is reconstructed under restricting the data collection procedure to promote the operation efficiency. To solve problems in scenes registration, a novel method for combining scenes in complex environments is proposed. The z-buffer structure from bird’s eye view is applied to estimate the moving path of the space shuttle car in an environment. The probability distribution of the position where we collect information from is consequently obtained. Aerial images are further used as global restriction and the base of modification. The success rate of scenes registration is 85%. A novel method for texture mapping is also established to overcome the problems of time-consuming and laborious while projecting panorama images to a model