整合環場影像及雷射掃描於大尺度虛擬實境模型之構建

Abstract

本研究整合環場影像及雷射掃描技術建構出大尺度之虛擬實境 (Virtual Reality, VR) 模型。相對於傳統虛擬實境之建構多為小尺度，本研究所發展的方法可構建大範圍空間尺度之虛擬實境模型—如擁有數十棟建築物之校園環境—且具有真實環境影像之貼圖材質。建構過程主要分為三個步驟：空間資訊擷取、三維模型建立及多場景接合。空間資訊擷取的過程中，藉由安排所獲取之資訊，可降低模型建立的複雜度，達成近即時運算之場景建模系統；單一場景資訊擷取所需時間約為12分鐘。三維虛擬模型的建立的步驟係利用空間資訊排列之規則將模型網格化，並配合Delaunay Triangulation 最佳化網格點並減量至60%，再利用環場影像於邊界條件上與模型具一致性之特點，簡化影像材質之對應。在多場景接合的方法上，本研究發展一套全域性接合方法，將三維空間資料簡化為二維資訊，並進行資料點擴展及型態分析，以提高接合正確率。以此接合方法於454.9公尺之長距離進行地圖重建實驗，誤差為0.1%。接合正確率與空間環境複雜度相關，實驗於具有規則特徵的環境，如具有規則牆面之臺大體育館，在平均距離46.7公尺，標準差8.9公尺的條件下，接合正確率可達100%。而在較複雜之環境，如臺大椰林大道之多植栽環境，在平均距離44.9公尺，標準差5.8公尺條件下，接合正確率則降為72％。若應用此接合方法於長寬十數公尺之室內環境，亦能正確接合。相對於傳統之迭代最近點算法 (Iterative Closest Point, ICP)，此接合法不需初始估測值，且在5分鐘內可完成計算，若給予估測值則約能在10秒內完成。本研究所發展的方法將可應用於模型重建、環境導覽及都市開發規劃等諸多應用。

This research proposes a method to reconstruct real-world objects in large-scale virtual reality by integrating the information from panoramic images and data acquired with a laser range sensor. Compared with most virtual reality applications, the proposed method is capable of building large-scale virtual reality models such as a campus environment including many buildings, and the reconstructed building's appearance is photorealistic. The procedure for building the virtual reality models mainly consists of three phases: the acquisition of the space information, reconstruction of 3D model, and integration of multiple scenes to form a large-scale virtual reality model. In the first phase of data acquisition, the complexity of the model reconstruction is reduced by properly arranging the acquired point cloud data. Therefore, the reconstruction of the 3D model is nearly real time. The time required to acquire the 3D point cloud data for a single scene was about 12 minutes. The reconstruction of 3D virtual reality model utilizes the well organized collected spatial data for meshing which also applies Delaunay triangulation for a data reduction to 60%. The texture mapping process is also simplified based on the correspondence of boundary conditions in the panorama image and the 3D model. Moreover, a global registration method was developed. It initially coverts the 3D spatial data into 2D information and then applies spatial propagation and shape analysis filters for enhancing the matching probability. When applying this method on a reconstruction of a map 454.9 meters in length, the error was 0.1%. The matching probability is affected by the complexity of environment to be reconstructed. When applying the method on an environment with regular features, such as the NTU gymnasium with mostly regular wall surfaces and with an average scanning distance of 46.8m and standard deviation of 8.8m, the matching probability was 100%. In a separate experiment for a more complex environment such as the NTU Royal Palm Blvd. with many trees and irregular structures, the matching probability decreases to 72%. The method was also successfully applied for the reconstruction of an indoor environment which is ten meters in width and height. Comparing with the iterative closest point (ICP) method, our approach has the advantage that it does not require initial guess for the registration of multiple scenes. The registration process can be accomplished in 5 minutes. Moreover, it can be completed in 10 seconds if the initial guess is given. The developed method can be used in various applications such as model reconstruction, tour guiding, and the city planning.