利用可攜雙全像式攝影機組建構人可到達環境影像的一種新式場景取像及瀏覽系統

Abstract

現今發展的各種街景系統，如Google Map及Google Earth，並無法取得某些特殊環境的場景影像，例如室內的階梯、陡峭的高山、花園的小徑等。本論文提出了一個利用可攜雙全像式攝影機組來取像及建構場景的瀏覽系統。本研究設計了一個輕巧的取像設備，能夠背負至人可到達的任何環境進行取像。取像設備包括兩個同軸且背對背的全像式攝影機(omni-camera)，可用以同時取得上下方的全像式影像(omni-image)，來涵蓋上下半圓的環境視野。 更詳言之，本研究首先將上下兩張全像式影像轉換成全景影像(panoramic image)，接著在影像的重疊區域上，利用全像式取像和特徵點幾何性質的限制來結合動態規劃(dynamic programming)技術，將上下兩張全景影像縫合成一完整的影像。為了能夠在場景中顯示不同的視角，本研究也藉由一透視對應表(perspective mapping table)來產生各種不同方向的透視影像(perspective-view image)，並縫合上下兩張透視影像來呈現一完整的影像。 另一方面，在場景瀏覽系統中，本研究提供六種方法讓使用者改變瀏覽的場景以及視角。在透視影像或路徑三視圖中，藉由點擊行走路徑上的方向箭頭或路徑三視圖中的任一位置點，系統會顯示對應場景的透視影像，同樣地，藉由滑鼠點擊全景影像或滑鼠拖移透視影像，系統會顯示其對應視角的場景影像。 本研究設計的系統適合用於取得任何環境的場景資訊，並能成功建構全景及透視影像，而且能讓使用者藉由不同的方式簡單地操作。實驗結果顯示所提出的系統在實際應用上確實可行。

As an improvement on conventional street view systems like the Google Map and the Google Earth, which cannot cover some special environment areas (like indoor space with stairways, mountains with steep roads, gardens or parks with narrow tracks, etc.), a scene imaging and browsing system using a portable two-camera omni-imaging device is proposed. The system is light and can be carried by a person on his/her back. The omni-imaging device is composed of a pair of omni-cameras aligned coaxially and back to back, which can be used to take an upper omni-image and a lower omni-image simultaneously. The two images cover respectively the upper and lower hemi-spherical fields of view of the environment with an overlapping image band. The omni-images are first transformed into two panoramic images. Then, the overlapping portion in them are utilized to stitch them together into an integrated panoramic image by a dynamic programming technique using some constraints about omni-imaging and feature point geometry in the overlapping image portion. The resulting larger panoramic image is used as a 360o map for traversing the entire environment via an interface. Then, for displaying different viewpoints at each scene spot in an environment, integrated perspective-view images from multiple view directions are constructed by stitching two perspective-view images generated from a perspective mapping table. Moreover, the scene browsing system provides six ways for a user to change the scene spot or the view direction. By clicking the arrows on a walking path in a perspective-view image or on a point in a three-view diagram, the perspective-view image of the corresponding scene spot can be displayed. Also, by clicking on a panoramic image or dragging a perspective-view image, the view direction of the displayed image can be changed to a corresponding one. The experimental results show that the designed scene imaging device is suitable for collecting scene information anywhere. Also, constructions of panoramic and perspective-view images using the two-camera device were also seen to be successful. Finally, the scene browsing system can be manipulated easily to browse the scene images in multiple ways. In short, the experimental results show the feasibility of the proposed scene imaging and browsing system for real applications.