利用多個兩層的Kinects做人體全身立體建構

Abstract

隨著3D掃描技術的進步，出現許多基於3D掃描器或成像系統的應用研究，其中之一是人物建模。大多數現有的建模方法耗費非常多的時間在圖像掃描以及後續的建模工程，所以本研究提出了一個兩層的快速建模系統，可在容忍時間內建造人體全身模型。這個快速三維建模系統是由十二台Kinects組成，六台放置在上層，其他則放置在下層。在上層中，有一台是Kinect v2，用來感測人臉，以提高所建構模型的臉部品質。所有的Kinects圍繞成一個圓圈，以拍攝站在三維建模系統中的人物。而這個三維建模方法最後可以利用三維列印機(3D printer)列印出所建人體全身模型。 本研究首先提出一方法來對準光軸和校準每台Kinect的平移和傾斜角度。接著，採用另一方法來學習每兩台相鄰Kinects間的幾何關係，該法是以距離加權相關測度（DWC）來進行3D圖像匹配。本研究還提出了一個快速的三維加權相關測度(3D DWC)比對方法來加速這個學習過程，該法是由一個計算二維加權相關測度(2D DWC)的快速演算法所衍生出來的3D版本。最後本研究根據校準幾何關係的參數，疊合建造3D圖像，以獲得三維人體全身模型。 此外，每台Kinect所得到的原始深度圖像通常有各種雜訊，會在系統的學習和模型構建的過程中造成問題。本研究使用了一些方法以降低或平滑這些雜訊，其中之一是用Poisson表面重建方法，以建構人體模型的網格(mesh)。在此重建方法中本研究使用不同的參數值來改變所建造模型的品質，並以系統化調整參數的方式去除雜訊，以得到平滑的建模結果。 上述方法的實驗結果皆甚良好，顯示本研究所提系統確實可行。

With the advance of 3D scanning technology, a lot of applications based on the use of 3D scanners or imaging systems have been investigated, and one of them is human modeling. Most of the existing modeling methods spend very much time in image scanning as well as in the subsequent works of model construction. So, a fast two-level 3D imaging system accompanied by a 3D whole human-body modeling method is proposed, which can be used to construct whole-body human models in an endurable time period. This fast 3D imaging system is composed of twelve Kinect devices with six of them being placed on the upper level and the others on the lower level. On the upper level, one of the Kinect devices is the Kinect version 2 which is used to sense particularly the human face to improve the quality of the face part of the constructed model. All the Kinect devices are arranged to form a circle to take the 3D images of a human standing in the middle of the system. And the 3D modeling method can be used to construct whole human models which can finally be printed out by a 3D printer. For the proposed imaging system, at first a method for aligning the optical axis and calibrating the pan and tilt angles of each Kinect device is proposed. Next, a method for learning the geometric relationship between every two neighboring Kinect devices is adopted, which is based on a 3D image matching technique using the distance-weighted correlation (DWC) measure. Also proposed is a fast 3D DWC method for speeding up this learning process, which is a 3D version derived in this study of a fast algorithm for computing the 2D DWC measure. Finally, according to the calibrated geometric relation parameters, the constructed 3D images can be merged to obtain a 3D whole human-body model. The original depth image acquired from each Kinect device usually includes many types of noise that may cause problems in the system learning and model construction processes. Several schemes are adopted to reduce or smoothing such noise. One of them is the Poisson surface reconstruction method, which is adopted to construct the meshes of the constructed human model. The different parameter values used in the reconstruction method influence the quality of the constructed model. A systematic adjustment of the parameter values to remove the noise to get a smoother modeling result is conducted in this study. Good experimental results are also shown to prove the feasibility of the proposed methods for real applications.