基于非线性优化的VR场景拼接方法研究

Research on VR Scene Stitching Based on Nonlinear Optimization

在虚拟现实(VR)技术中,为获得更宽的视场需要将多幅图片通过拼接构建出全景.目前大多数基于特征的图像拼接算法采用单应性矩阵变换或保持内容的变形来实现图像对准.基于单应性矩阵变换方法很难处理视差,而保持内容变形方法不能保持图像的结构特性.本文提出了一种精确的单应性矩阵的求解方法,由匹配点和特征点求得全局单应性矩阵,并通过非线性优化方法对该结果进行优化.进一步针对局部变形带来的误差,结合几何和光度约束来设计代价函数,通过最小化该函数获得更好的图像对准结果,从而产生精确的拼接图像.通过各种开放数据集的实验结果表明,本文提出的图像配准方法效果最好.

In virtual reality(VR),in order to obtain a wider field of view,multiple images need to be stitched to construct a panorama.Despite significant advances in image stitching in recent years,this problem still lacks a reliable solution.Most feature-based image stitching algorithms perform image alignment based on homography matrix transformations or by maintaining distortion of the content.It is dicult to deal with parallax based on the homography matrix transformation method,and the content modification algorithm cannot maintain the structural characteristics of the image.In this paper,an accurate method for solving the homography matrix is proposed.The global homography matrix is obtained from the matching points and the feature points,and the result is optimized by the nonlinear optimization method.Further errors caused by local deformation,combined with geometric and photometric constraints to design our cost function,minimize this function,and obtain better image alignment,resulting in accurate image stitching.Experimental results through various open datasets show that our proposed method is superior to the most advanced image spelling algorithm.