自兩幅影像中估測相機位置參數

Abstract

立體數位內容在多媒體領域具有廣泛應用的潛力及市場需求，新型態的技術也因應而生。像是自由視點視訊(FTV, Free Viewpoint Television)，是以任意視點合成技術為核心。許多合成演算法是利用多重影像以及其對應的深度資訊圖來產生虛擬試點的影像以達到任意視點的效果。

我們利用兩組輸入影片，分別為不同的彩色攝影機以及相對應的深度攝影機所拍攝，試著結合這些輸入來產生新的立體場景。此立體場景的構成是由其中一組影片輸入背景，另一組影片輸入前景物體。因此，這兩組影片中的不匹配，將會影響合成之後個影像品質。在此論文中，我們主要將討論攝影機參數以及攝影機運動的不匹配，在兩組影片中，調整背景場景以達成背景置換的目的，並使合成影像符合主觀視覺

3D video becomes a popular topic recently; video composition is one of the challenge applications. Given two separate videos captured by two sets of multiple cameras, we combine them to form a new virtual video, which contains the foreground objects from one video and the background objects from the other. However, there are often some mismatches need to adjust due to the different capturing devices and their relative poses and movement. In this thesis, we focus on the camera movement mismatch of foreground and background videos. If the camera motions of these two sequences differ from each other, the virtual composite may look unnatural. We propose two algorithms to estimate the camera (relative) extrinsic parameters. Then, based the obtained parameters, we create a virtual background video to match the camera movement of the foreground video (objects). After the adjustment, the view synthesis software VSRS is used for image composition frame by frame, and the virtual stereo video can thus be generated.