在當前科技發展的背景下，由於虛擬現實(VR)、3D列印等技術的蓬勃發展，對建立3D模型的需求也隨之增加。然而傳統的3D建模方法不僅須具備專業知識，還需要專業的設備和軟體，既耗時又昂貴。而本研究提出的方法旨在降低3D模型建立的時間和金錢成本，並降低使用門檻。使用Kinect V2或其他類似能夠同時獲取RGB影像及深度資訊的設備，即可使用本研究的方法，實現目標物體的3D點雲生成，方便進行後續的3D建模等相關應用。
本研究旨在基於多張RGB-D影像進行3D點雲重建，使用的設備是微軟的Kinect V2感應器。首先，對目標物體進行環狀拍攝，獲取其RGB影像及深度資訊。將這些資訊經過畸變校正處理，將彩色影像與深度影像進行對齊，從而生成初始的點雲數據。點雲數據通常會包含許多噪聲和不完整的區域，因此需要進行進一步處理。在點雲前處理階段，首先選取感興趣區域(ROI)，以確保只處理目標物體的相關數據。此階段包含了使用RANSAC去除點雲中包含地板的部分，及使用離群值去除如基於密度的空間分群演算法(DBSCAN)清除點雲中的噪點並將點雲分群。此外，對點雲進行下採樣處理以減少數據量，提高後續的處理速度。隨後，使用迭代最近點演算法(ICP)進行點雲對齊。ICP演算法在計算兩個點雲之間的最佳旋轉及平移矩陣，分為點對點和點對面的對齊方法。這一過程需要經過多次迭代以達到精確對齊，最終生成一個目標物的完整點雲。

With the rapid development of technologies like virtual reality (VR) and 3D printing, the demand for 3D model creation has increased. Traditional 3D modeling methods are time-consuming, expensive, and require specialized knowledge and equipment. This study proposes a method to reduce the cost and time of 3D model creation, using Kinect V2 or similar devices that capture both RGB images and depth information. This method facilitates the generation of 3D pointclouds, aiding in subsequent 3D modeling applications.
The study focuses on 3D model reconstruction based on multiple RGB-D images using Microsoft's Kinect V2 sensor. The target object is captured in a circular manner to obtain its RGB images and depth information. These data are processed through distortion correction and alignment of color images with depth images to generate initial pointcloud data. The pointcloud is then preprocessed by selecting the region of interest (ROI) and removing noise and outliers using methods like RANSAC and DBSCAN. Downsampling is performed to reduce data volume and improve processing speed. The Iterative Closest Point (ICP) algorithm is used for pointcloud alignment, calculating the optimal rotation and translation matrices between pointclouds. Multiple iterations ensure precise alignment, resulting in a complete pointcloud of the target object.

論文審定書 i
致謝 ii
摘要 iii
Abstract iv
目錄 v
圖目錄 vii
表目錄 x
第一章 緒論 1
1.1 動機與目的 1
1.2 文獻回顧 2
1.3 研究方法與步驟 6
1.4 論文架構 6
第二章 點雲 8
2.1 深度圖與點雲介紹 8
2.2 針孔相機模型(Pinhole camera model) 10
2.3 深度感測方法 11
2.3.1 雙目視覺(Stereo vision) 11
2.3.2 結構光(Structured light) 13
2.3.3 飛時測距(Time of flight) 14
第三章 微軟Kinect V2感應器 16
3.1 Kinect V2規格與介紹 16
3.2 相機失真校正(Camera calibration) 17
3.2.1 電腦視覺中的齊次座標 17
3.2.2 相機內部參數與外部參數 17
3.2.3 畸變校正 18
3.2.4 校正步驟與結果 20
3.3 彩色影像與深度影像對齊 22
3.3.1 座標轉換 22
3.3.2 對齊步驟與結果 24
3.4 從Kinect V2取得點雲步驟與結果 27
第四章 點雲前處理 28
4.1 點雲ROI選取與離群值去除 28
4.1.1 找出點雲中的地板並移除 29
4.1.2 點雲聚類 33
4.2 點雲下採樣(Downsampling) 36
第五章 點雲對齊 39
5.1 點對點的ICP 39
5.2 點對面的ICP 49
第六章 實驗流程與結果 52
6.1 實驗流程 54
6.2 實驗結果 62
6.2.1 Redwood-3dscan Dataset 的實驗結果 63
6.2.2 Kinect V2 自有數據集的實驗結果 66
6.2.3 比較與分析 68
第七章 結論與未來展望 69
7.1 結論 69
7.2 未來展望 69
參考文獻 71

[1] Kim, C., Fanello, S. R., Rhemann, C., Izadi, S., Hilliges, O., & Davison, A. J. "Color and Depth Image Correspondence for Kinect v2," In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2015, pp. 36-43.
[2] Choi, S., Zhou, Q.-Y., & Koltun, V. "A Large Dataset of Object Scans," In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2016, pp. 16-24.
[3] Blender Foundation. "Animation Example," [Image]. Retrieved from https://www.blender.org/wp-content/uploads/2019/07/animation01-1280x720.jpg?x12104. 2019.
[4] Merigot, Q., & Oudet, E. "Analyzing the Squared Distance-to-Measure Gradient Flow System with k-Order Voronoi Diagrams," Journal of Computational Geometry, 3(1), 2012, pp. 1-23.
[5] 周錫珉. "具有空間三維幾何量測之八進位井字編碼結構光的設計, "國立中央大學光電科學與工程學系, 碩士論文. 2018.
[6] Wikipedia contributors. "Time-of-flight camera illustration," [Image]. Retrieved from https://en.wikipedia.org/wiki/Time-of-flight_camera#/media/File:20200501_Time_of_flight.svg. 2020.
[7] Googleusercontent. "Kinect v2 Sensor," [Image]. Retrieved from https://lh6.googleusercontent.com/-lpQGeSd3uDE/VKDgQIx5WsI/AAAAAAAAxY4/FwvfoKo8jZk/s640/k4wv2-sensor.jpg. n.d.
[8] Bradski, G., & Kaehler, A. "Learning OpenCV: Computer Vision with the OpenCV Library," O'Reilly Media, Inc. 2008.
[9] "Matlab Document: What is Camera Calibration," [Online]. Available: https://www.mathworks.com/help/vision/ug/camera-calibration.html. n.d.
[10] Núñez, P., Vásquez, A., & Tejera, G. "Calibration of Kinect for Xbox One and Comparison between the Two Generations of Microsoft Sensors," Sensors, https://doi.org/10.3390/s141121940, 14(11), 2014, pp. 21940-21957.
[11] Rothwell, C., Forsyth, D.A., Zisserman, A., & Mundy, J.L. "Extracting Projective Structure from Single Perspective Views of 3D Point Sets," In Proceedings of the Fourth International Conference on Computer Vision IEEE Press, 1993, pp. 573-582.
[12] Fischler, M. A., & Bolles, R. C. "Random Sample Consensus: A Paradigm for Model Fitting with Applications to Image Analysis and Automated Cartography," Communications of the ACM 24(6), 1981, pp. 381-395.
[13] Ester, M., Kriegel, H. P., Sander, J., & Xu, X. "A Density-based Algorithm for Discovering Clusters in Large Spatial Databases with Noise," In Proceedings of the 2nd International Conference on Knowledge Discovery and Data Mining (KDD) , 1996, pp. 226-231.
[14] Wikipedia. "DBSCAN – Illustration," Wikimedia Commons, https://zh.wikipedia.org/wiki/DBSCAN#/media/File.svg. Accessed August 2024.
[15] "Levenberg-Marquardt Algorithm Visualization," Miro, Medium. Available at: https://miro.medium.com/v2/resize:fit:1350/format/0*o_BoNtBbwvAiNFLB.gif. Accessed August 8, 2024.
[16] Besl, P.J., & McKay, N.D. "A Method for Registration of 3-D Shapes," IEEE Transactions on Pattern Analysis and Machine Intelligence 14(2), 1992, pp. 239-256.
[17] Chen, Y., & Medioni, G. "Object Modelling by Registration of Multiple Range Images," Image and Vision Computing, 1992, 10(3), pp. 145-155.