Fast 3D Human Modeling from Outside-In Views Using Multiple KINECT Devices
|關鍵字:||3D掃描;深度感測器;人物建模;校正;3D scanning;depth-sensor;human modeling;calibration|
基於針孔相機原理，本研究首先提出一個座標轉換系統，可將KINECT得到的深度影像轉換為一3D空間中的點集合，藉以將KINECT得到的彩色影像及深度影像組合成一3D影像。此外，本研究亦對3D影像的彎曲現象做幾何校正。接著，本研究使用一「考慮色彩資訊的距離加權相關度(distance-weighted correlation, DWC)」對3D影像作比對，來校正相鄰KINECT的關係參數。另外，本研究也提出一方法，藉以使用校正後所得到的參數，來整合3D影像成為3D人物模型。
With the advance of 3D scanning technology, there are lots of applications based on the use of 3D scanners. Recently, a very popular application of 3D scanning is human modeling. But most of the existing modeling methods spend very much time in image scanning as well as in the subsequent works of system calibration and model construction. So, in this study, a fast 3D imaging system accompanied by a 3D modeling method is proposed. The fast 3D imaging system is composed of six KINECT devices, which together are arranged to form a circle to take KINECT images of a human standing in the middle. And the 3D modeling method can be used to construct human models which can finally be printed out by a 3D printer. For the proposed imaging system, at first based on the pinhole camera model, a method is employed to construct 3D images from the color and depth images acquired with a KINECT device. Also, a scheme is employed to correct the bending phenomenon existing in the constructed 3D image by geometric correction. Next, a method for calibrating the relationship between every two neighboring KINECT image frames is proposed, which is based on 3D image matching using the distance-weighted correlation (DWC) measure as well as the color information of the KINECT color images. Also proposed is a method for merging the constructed 3D images according to the calibrated relation parameters, resulting in a 3D human model. The original depth image acquired from the KINECT device usually includes many types of noise that may cause problems in the system calibration and model construction processes. A method for dealing with such noise before constructing the final human model is therefore proposed, which applies various types of filtering to each depth image to get a smoother and noise-reducing result. Also proposed is a method for finding an optimal KINECT color image for use in assigning colors to the surface pixels of the constructed model. The constructed human model is printed out finally with a 3D printer. Good experimental results are also shown to prove the feasibility of the proposed methods for real applications.