使用彩色深度感測器之物體模型重建系統與高解析度紋理貼圖

Abstract

三維物體模型重建在許多領域中都扮演著非常重要的角色，像是3D 列印技術、逆向工程以及醫療領域中所使用到的義肢與矯正器等等。隨著近幾年3D 列印技術的蓬勃發展，如何有效率地掃描現實生活中的環境與物件變得更為重要。本研究的目的是發展出一套自動化的物體模型重建系統，掃描重建的結果為一個精緻的虛擬模型，同時，亦能夠呈現高解析度的紋理貼圖於模型表面上。

我們的掃描裝置是使用結構光源原理來量測待測區域的深度值，本系統主要分為三個部分，第一部分是相機校正與對位，第二部分為資料取得的方法與點雲的對位，第三則是網格模型的重建與紋理貼圖。我們使用張正友校正法計算出相機參數，該方法中所使用的棋盤格校正法亦同時運用在多個相機視點的對位轉換上。在取得待測區域的深度圖之後，我們使用背景去除與去雜訊技術，以獲得各個相機視點的前景點雲。接著我們將所有點雲由相機座標系統轉換至世界座標系統，並透過加入顏色資訊做為參考的迭代最近點演算法以及全域最佳化技術來精進點雲的對位關係。最後，我們使用波森表面重建演算法重建模型網格，並同時使用相機的內、外部參數模擬相機拍攝時的相機視點，以利我們將高解析度紋理貼圖貼合至模型表面之上。

根據我們的實驗結果，此系統非常地有效率，只需花費兩分鐘左右的時間就能夠獲得最終的掃描重建成果，其中包含了掃描程序約為106 秒，以及運算程序所需時間為22 秒。另一方面，在點雲對位演算法的強化之下，能夠使最後的模型成品與其紋理貼圖極為精緻，模型的尺寸誤差值約為4.25 個百分比。此外，我們的系統提供全自動化的操作程序，一鍵掃描即可產出物體模型重建的成果，達到高度自動化的目的。

3D model reconstruction is essential to 3D printing, reverse engineering, prototyping, orthotics and prosthetics. Nowadays, along with the rapid growth in 3D printer, how to scan a real-world object efficiently becomes more important. We developed an automated object model reconstruction system, which produced a delicate mesh model with high-resolution texture.

In the proposed system, we used structured light device as our scanning equipment. Our work contains three main parts. The first part is camera calibration and coordinate alignment. The second part is data acquisition and point cloud alignment. The final part is model reconstruction and texture mapping. In the procedure of calibration, Zhang’s method was used to calibrate our camera. The chessboard approach helped us with finding the coordinate system alignment between multiple views. Next, we used background subtraction and radius outlier removal method to obtain a group of foreground point cloud. We then transformed all point clouds from the camera coordinate system to the world coordinate system by calibration results. With a group of noise-free clouds, we used the color-supported ICP algorithm to get an accurate alignment result. We further refined the relation between camera poses by a global optimization method. The final part was the reconstruction of mesh model and texture mapping. We applied Poisson mesh reconstruction technique, constructed a mesh model from the point cloud. With the camera intrinsic and extrinsic parameters, we simulated the camera model, and back projected high-resolution photos on the model in the end.

According to our experimental results, the proposed system can achieve efficient object scan. The whole scanning and reconstruction process can be completed in about 2 minutes, including 106 seconds of the scanning time and 22 seconds of the computing time. On the other hand, with the camera calibration, alignment algorithm improvement and optimization, our approach can accurately reconstruct the 3D object model. The model size is about 4.25% of the real object size. Besides, our system provided an automated operating interface. It is able to complete an entire scanning procedure in one-click.