由視訊中自動建立環場接合圖之研究

傳統上，在建立mosaic影像時，在拍攝階段皆限制於平面的場景或物體，亦或者僅能將取像設備固定在某一位置做旋轉拍攝。這些限制下的影像接合圖（mosaic）在先前的研究中已有不錯的研究結果，本研究則是基於S. Peleg的Manifold投影概念來建立環場影像接合圖。運用此概念來建立環場影像接合圖時，允許在拍攝的同時具有平移運動以及旋轉運動，因而限制較少也較符合實際在拍攝時的攝影機運動狀態，此外，在找尋鄰圖框間的相關運動估測時，本研究利用Winner-Update的快速區塊比對演算法，使得本研究得以即時（real-time）的方式邁立環境影像接合圖。

Traditionally, researches on building panoramic mosaics have been restricted to planar objects on scenes or assume the camera motion to be pure rotation. That is, camera is rotating around its optical center. Restricted to these two cases, previous related researches have presented good experimental results to build panoramic mosaic. However, more general camera motions should consist of translation and rotation together. In this paper, we use the concept of manifold projection proposed by S. Peleg et al.,[8][9] to create mosaics from videos. This projection model enables us to project images onto manifold mosaic under a more general camera motion condition, such as motion caused by shooting with a hand-held camera. On the other hand, this model does not require the scene to be planar. In this research, we also use the Winner-Update fast block matching algorithm [1] to accelerate the speed of motion estimation between two contiguous frames for the purpose of real-time construction of building panoramic mosaics.