利用深度感測器之三維物體模型重建與擬真互動

Abstract

在本研究中我們使用Microsoft Kinect深度感測器來發展三維物體模型重建與擬真人機互動系統，提供使用者自然觀看此物體的視覺互動體驗。使用Microsoft Kinect深度感測器的原因是因為它容易取得、售價便宜、可靠度高。 本系統主要分為三個部份，第一部份為前景偵測與點雲對位(registration)，第二部份是三維物體模型重建，第三部份是擬真人機互動應用。前景偵測主要使用背景相減法(background subtraction)，針對點雲對位的方法，我們使用改進迭代最近點(iterative closest point)演算法，基於配對目標的最近點，增加色彩相似度的比較。三維物體模型重建則先透過資料整合(data integration)和色彩校正(color correction)來保持點雲密度及校正不同點雲間的色彩差異。此外，我們使用波森表面重建(Poisson surface reconstruction)演算法來重建三維表面模型。在擬真人機互動中，本系統會先透過深度感測器不斷追蹤使用者的頭部座標，並計算出對應離軸透視投影(off-axis perspective projection)矩陣來變更每一視點之物體畫面，同時，使用者可以使用右手手勢來旋轉此三維物體。 本論文中，我們發展了三維物體重建與擬真人機互動系統，在建構三維物體模型後，搭配視覺化的擬真互動，營造栩栩如真的視覺互動體驗。根據實驗結果顯示，我們提出的改進迭代最近點演算法可改善對位的準確性。

In this study we developed a 3-D object reconstruction and immersive interaction system using Microsoft Kinect sensor. This system can construct a three-dimensional model of the object, and provide users with vivid visualization experience. The advantages of Microsoft Kinect sensor are their low prices and high accuracy. The system contains three parts. The first part is foreground segmentation and point clouds registration. The second part is model reconstruction of three-dimensional objects. The third one is immersive human-computer interaction system. Background subtraction algorithm is used to segment the foreground object from background scene. We proposed a the modified iterative closest point algorithm to estimate the transformation between adjacent point clouds. Color similarity of corresponding points is considered in this algorithm. Data integration and color correction are applied for maintaining point cloud density and removing of the color difference between corresponding point clouds. Then we apply Poisson surface reconstruction algorithm to reconstruct 3-D surface model. Furthermore, we estimate the off-axis perspective projection matrix from the user position captured by Kinect sensor. A simple gesture is used to rotate object model. In this work, we have developed a 3-D object reconstruction and immersive interaction system which can reconstruct a three-dimensional object model with immersive interaction experience. According to our experimental results, the modified iterative closest point algorithm is more reliable and accurate than the original iterative closest point algorithm.