Title: | 使用色彩與深度感測器之三維物體模型重建與紋理貼圖 3D Object Model Reconstruction with Texture Mapping using RGB-D Camera |
Authors: | 許志維 Hsu, Chih-Wei 陳永昇 Chen, Yong-Sheng 資訊科學與工程研究所 |
Keywords: | 模型重建;reconstruction;kinect |
Issue Date: | 2012 |
Abstract: | 在電腦視覺與逆向工程領域中,準確的建立三維物體模型是一項值得研究的 課題,在本研究中我們使用Microsoft Kinect for Windows 做為深度與影像感測器 來發展物體的三維模型重建與紋理貼圖系統,這種深度感測器支援近距離模式, 可以支援到距離物體40 公分的距離,並且容易取得、可靠度高。 本研究方法主要分為三個部分,第一部分為前景擷取,第二部分為不同視點 間的點雲對位(registration),全域的整合(global refinement)與三維物體模型重建 (reconstruction) , 最後一個部分為不同影像對三維物體的紋理貼圖(texture mapping)。前景擷取主要使用背景相減法找出前景點雲並移除雜訊以增加之後對 位的準確性,並環繞物體拍攝14 張影像,包括物體的頂部與底部。第二部分點 雲的對位使用點到平面的迭代最近點(point-to-plane ICP)演算法來進行資料的對 位,三維物體模型重建的部份則使用波森表面重建(Poisson surface reconstruction) 演算法來重建三維表面模型。在第三部分,我們改善由不同影像紋理貼圖到同一 物體所產生的不連續感,整合不同視角所拍的影像,完成三維物體的紋理貼圖。 本論文中,我們發展了物體的三維模型重建與紋理貼圖系統,根據實驗結果 顯示,使用固定的初始矩陣可以改善對位的速度,而我們提出的紋理貼圖方法可 以改善不同影像間紋理貼圖的效果,增加物體紋理貼圖之後的精細度,呈現出擬 真的三維物體模型。 In the computer vision and reverse engineering field, accurate reconstruction of 3D models is an essential topic. In this study, we developed a system for whole 3D object model reconstruction and texture mapping by using Microsoft Kinect for Windows. The depth sensor of Microsoft Kinect for Windows supports near mode which enables the depth camera to scan in front of an object as close as 40 centimeters without missing data. Due to high accuracy and low prices of Microsoft Kinect for Windows, we selected Kinect as our system development device. The system contains three parts. The first part is foreground acquisition. The second part is registration of multiple views point cloud and 3D object surface reconstruction. The last part is multiple views texture mapping. Foreground acquisition uses background subtraction to obtain foreground point cloud and remove outlier points. We took 14 images from an object by a stationary Kinect-- including the top and bottom-- and used the point-to-plane iterative closest point (ICP) algorithm to estimate the transformation in adjacent point clouds. Once we got integrated final point cloud, we applied the Poisson surface reconstruction algorithm to reconstruct a 3D surface model. Then we used multiple images to map onto a 3D surface model and improved the quality of texture mapping by reducing the boundaries triangles that mapped from different images. In this study, we have developed whole 3D object model reconstruction and texture mapping system. According to our experimental results, using a predefined initial matrix for initial alignment can accelerate the registration part and the multiple views texture mapping method improves the quality of the texture mapping, which makes the reconstructed 3D object model more realistic. |
URI: | http://140.113.39.130/cdrfb3/record/nctu/#GT070056070 http://hdl.handle.net/11536/71840 |
Appears in Collections: | Thesis |