中文摘要：

全景视图在视频会议、虚拟现实等领域有广阔应用前景，而现有的合成技术对大角度旋转等失配以及存在视差的情况效果不佳，为此提出一种鲁棒性较好的全景视图生成算法．该方法首先利用可重复性特征估计透视运动模型，从而对相邻视图进行配准；然后利用显著特征对视图分层，并对重叠区域进行非线性融合，以减小视差对重叠场号的影响．多幅不同条件视图的合成实验表明，该算法在大角度旋转及视差等情况下仍能较好地合成视图．

英文摘要：

Panorama has various application prospects such as teleconference, virtual reality, etc. However, in the case of large misregistration or parallax, many existing algorithms perform poorly or even fail to work through. To tackle this challenging situation, a robust approach was proposed to generating panorama. This method first estimated the perspective motion model to align the different views by repeatable features. And then it ordered the images into different layers based on salient features before blending. The experimental results show that the proposed method can effectively synthesize different views with misregistration or parallax.