以多個兩層式的Kinect系統做彩色人體全身模型之快速建構

Abstract

現今，3D列印這項技術變得越來越熱門，它是一項使用三維印表機(3D printer)的科技，能以適當的材料列印出多樣的模型。而另一項技術─三維掃描(3D scanning)也是不可或缺的，這項技術使用三維掃描器取得物體的顏色以及深度資訊。在本研究中，使用了這兩項技術來建構並列印人體模型。 做為邱與蔡[2]研究的延續，本研究使用其所設計的「兩層式多Kinect系統」(由一台第二代Kinect以及十一台第一代Kinect所組成)來掃描人體的色彩以及深度資料。主要的目標是要建構出一個平滑的彩色全身模型，供3D列印之用。 本系統程序包含兩部分，即「學習」與「建構」。在學習的過程中，起初會以一種使用到膚色偵測的方式來分離出頭部。接著再運用「顏色過濾距離權重相關係數(color-filtered distance-weighted correlation, CFDWC)」，來校正人體三個部分(頭部、上半身以及下半身)的疊合參數，並且利用「距離映射(distance map)」加速校正的過程。 在建構的過程中，本研究將校正完的疊合參數應用在Kinect系統得到的三維影像上，產生出三維人體全身資料。為了建構平滑的全身模型，起初先將全身資料，利用邊緣偵測和膚色偵測的方法，分割成三部分─頭部、中段身體以及腳。接著再對不同的部位，使用不同的建構參數，來建構出平滑程度相異的模型。藉由組合這些模型，即可得到一個平滑的全身模型。 接著基於「尋找數個最近點」的方式(k nearest neighbor, kNN)，本研究將原本的單色模型上色，建構出彩色的人體全身模型。最後，在進行3D列印前，再對人體模型的CMYK色彩進行調色，讓列印的結果較接近螢幕上所看到的色彩。 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。

Nowadays, 3D printing has become more and more popular. It is a technology which can be used to print various models in proper materials with a 3D printer. And another related technology, 3D scanning, is also considered indispensable, which can be employed to capture the color and depth information of an object by using a 3D scanner. In this study, both technologies are used to construct and print human models. As a continuation of the research of Chiu and Tsai [2], their two-level multi-Kinects system, which is composed of one Kinect version 2 device and eleven Kinect version 1 devices, is used in this study to scan the color and depth data of a human body. The major goal is to refine their system to construct smooth colored whole-body models for 3D printing. The processes of the proposed system include two phases, learning and construction. In the learning process, at first the human head part is segmented out by a method based on skin-color detection. Then, the merging parameters of the 3D data of three human-body parts, namely, the head, the upper body, and the lower one, are derived by a newly-proposed color-filtered distance-weighted correlation (DWC) measure together with a speedup measure based on the use of distance map. In the model construction process, the calibrated merging parameters are applied to the 3D images obtained through vision-based transformations from the color and depth data of the Kinect devices, yielding a set of well-merged 3D whole-body data. To construct a smooth whole-body model, at first the 3D whole-body data are split into three parts, the head, the middle body, and the legs, by edge detection and skin-color segmentation techniques. Then, different modeling parameters on the three parts are applied respectively to construct partial models with different degrees of smoothness for the three parts. By composing these partial models of the three parts, a smooth whole-body model is acquired. To construct a colored human-body model, a coloring method based on the concept of k nearest neighbor (kNN) is proposed to assign colors to the vertex points of the original monochrome model. Finally, before taking the constructed colored model to the 3D printer, the colors of the model in the CMYK color space are adjusted to make the colors of the to-be-printed result visually close to those of the model constructed originally and seen on the screen. Good experimental results are also presented to show the feasibility of the proposed methods and the system for real applications of 3D human-body scanning and printing.