三維重建技術應用於環場影像之研究

The three-dimensional reconstruction technology apply to panorama

朱榮堂(Jung-Tang Chu)

指導教授 ： 孫天龍博士

元智大學/工程學院/工業工程與管理學系/碩士(2001年)

摘要

一個3D虛擬場景可以使用幾何式的方法或影像式的方法來描繪出來。幾何式的3D場景首先需要在場景中建構3D電腦圖學的模型，其優點為使用者能在此場景自由遊走及從任意的視點及視角來察看此場景。此外，因為全部的幾何模型都可以在電腦上製作出來，所以此場景的建構可以全部是“虛擬”的。但是，製作複雜場景的幾何模型是非常困難且費時的，而且其執行的速度會隨著幾何模型複雜度的增加而降低。另一方面，影像式的3D場景較容易製作，且其執行速度與場景的複雜度是互相獨立的。但是由於場景是由2D的影像所構成的，所以此場景中的遊走通常會受到限制。例如環場影像的方式，其只允許使用者站在一個固定的地方轉動其視角。在前進及後退的移動方面，其是用放大及縮小影像的方式來模擬此效果。由於此種方式並無法處理“遮蔽效應”，所以使得場景中的遊走並不逼真。
在本論文中，我們使用一種混合的方式來改善環場影像的遮蔽效應。3D場景的影像被區分為前景的物體及背景的影像。背景影像被用來建構環場影像；同時並使用相機自校準的透視重建技術[Fusiello, 2000; Zhang, 1998; Faugeras, 1992; 汪慶祥, 2000; 陳郁中, 2000]及歐幾里德重建技術[Devernay and Faugeras, 1996; 陳郁中, 2000]將前景物體上可以看到的3D幾何表面重建出來。當繪製此3D場景時，先繪製背景的影像，然後再繪製前景物體的3D表面。實驗結果證實此混合的繪製方式比純影像式的繪製方式能呈現出更好的遮蔽效應。

關鍵字

三維重建 ； 環場影像

並列摘要

A 3D virtual environment could be rendered using either a geometry-based approach or an image-based approach. Geometry-based 3D scene requires constructing the 3D computer graphic models in the scene first. It has the advantage that the user could be navigated the scene freely, examining the scene from any viewpoint and at any view direction. Also the scene built could be totally “virtual”, as the whole geometric models could be build on the computer. But generating the geometric models in a complicated scene is difficult and time-consuming. Moreover, the rendering speed is proportional to the complexity of the geometric model. The image-based 3D scene, on the other hand, is easier to generate and the rendering speed is independent of the scene complexity. But navigation is usually restricted because the scene is composed of 2D images. The panoramic view approach, for example, only allows the user to stand at a fixed position and turn his/her view directions around. For forward and backward movement, the image is enlarged or reduced to mimic the effect. Such approach could not handle the “overlapping effect”, hence making the scene navigation less realistic.
In this work we improve the “overlapping effect” in panoramic view by a hybrid rendering approach. The image of a 3D scene is divided into background image and foreground objects. The background images are used to construct the panoramic view; while the 3D geometry of the observable faces on the foreground objects are reconstructed using the perspective reconstruction technique for camera self-calibration[Fusiello, 2000; Zhang, 1998; Faugeras, 1992; Wang, 2000; Chen, 2000] and the Euclidean reconstruction technique[Devernay and Faugeras, 1996; Chen, 2000]. When rendering the 3D scene, the background images are drawn, followed by rendering the 3D faces of the foreground objects. Experimental results showed that this hybrid rendering approach presents much better overlapping effect than the pure image-based rendering approach.

並列關鍵字

3D reconstruction ； panorama

參考文獻

Bao, H., L. Chen, J. Ying and Q. Peng, “Nonlinear view Interpolation,” Proceedings of the Sixth Pacific Conference on Computer Graphics and Applications, pp. 61-69 and 225, Singapore, October 26-29, 1998.

Chan, Y. F., M. H. Fok, C. W. Fu, P. A. Heng and T. T. Wong, “A panoramic-based walkthrough system using real photos,” Proceedings of the Seventh Pacific Conference on Computer Graphics and Applications, pp. 231-240 and 328, Seoul, Korea, October 5-7, 1999.

Darsa, L., B. Costa and A. Varshney, “Walkthroughs of complex environments using image-based simplification,” Computers & Graphics, vol. 22, no. 1, pp. 55-69, February, 1998.

Fu, C. W., T. T. Wong and P. A. Heng, “Triangle-based view interpolation without depth buffering,” Journal of Graphics Tools, vol. 3, no. 4, pp. 13-31, 1998.

Fusiello, A., “Uncalibrated Euclidean reconstruction: a review,” Image and Vision Computing, vol. 18, pp. 555-563, 2000.