利用多部KINECT建構環車3D行車紀錄器及其應用

Abstract

本研究利用架設多台KINECT感測器在車子上來建立三維的取像系統做行車紀錄器之應用。每台KINECT感測器的位置都不盡相同，在最後使用了14台的KINECT感測器來環繞起車子週遭的360o。為了在計錄時更佳的快速，本研究使用了兩台電腦，並且每台電腦控制了7台的KINECT感測器。 為了建構出3D的行車紀錄器，第一步本研究提出了一個方法建立出環車的三維模型。這個方法是建立在針孔成像的模型之上。並且這些模型可以借由顯像在3維空間的技術，由各個不同的角度觀看.。 第二步，本研究提出了一個方法來校正兩台相鄰的KINECT感測器之間的關係，此法是利用到了最近點迭代法之演算法。此外利用預先學習的資訊，校正也可以更加快速。 第三步，本研究提出了一個方法來將車子上的多台KINECT感測器所得資料做整合。所用的方式就是將上一步中的校正所得的相對關係，並且作用在每對相鄰的KINECT感測器之後再做座標的轉換。 第四步，本研究提出了一個方法來將多張的環車KINECT感測器所得之彩色影像給接成一張，此法是建立在一個自動接圖的演算法之上。最後接出來的全景圖會成為3D行車紀錄器的背景，因為在室外能得到的三維資訊其實是不夠多的。 最後本研究將前述的三維模型，以及前一步中所述的背景圖整合。利用背景圖投影到球形之上，並且有近景的部分轉換到對應的角度，實驗的結果可以看到在近的部分有3D的模型，而在遠的部分則有背景的部分。 最後，上述方法的實驗結果良好，顯示出本研究所提系統確實可行。

In this study, a 3D imaging system is constructed for use as an event data recorder by affixing multiple KINECT devices to the body of a vehicle. The position of each KINECT device is different, and totally the views of the 14 KINECT devices in the system cover the 360o car surround. To speed up the data recording work, two computers are used, each controlling seven KINECT devices. For the proposed system, at first a method for creating a 3D model of around-car objects is proposed. This method is based on the pinhole camera model. The constructed around-car object model can be rendered in the 3D space, and seen from specific views. Secondly, a method for calibration of the relationship between neighboring KINECT devices is proposed, which is based on the iterative closest point (ICP) algorithm. Moreover, the calibration is speeded up by the use of some learned information. Then, a method is proposed for merging the images acquired from multiple KINECT devices around the car. With the relationship parameters acquired after calibration, the method transforms accordingly the coordinates between each pair of neighboring KINECT devices. Also, a method is proposed for panoramic image creation by stitching multiple color images into a single view. This stitching method is based on an automatic stitching algorithm. The panoramic image is used to act as a background in the displayed scene because the 3D image data constructed by the system sometimes are not enough. Finally, a method for combining 3D images with the 2D panoramic image is proposed. The panoramic image is projected onto a spherical shape which, together with the 3D around-car object model, is rendered for inspection of the geometric relation between the nearby object and the background scene. Good experimental results are also shown, which prove the feasibility of the proposed methods for real applications.