任意视角的多视图立体匹配系统  

Integrated multi-view stereo matching system with arbitrary viewpoints

作　　者：郭禾[1] 傅新元[1] 陈锋[1] 王宇新[1] 李寒[1] 

机构地区：[1]大连理工大学计算机科学与技术系,辽宁大连116024

出　　处：《计算机工程与设计》2007年第12期2873-2877,共5页Computer Engineering and Design

摘　　要：为了得到高精度深度图,通过特征点提取与匹配、计算基础矩阵F、引导互匹配、捆集调整等一系列技术,求出每幅视图在同一空间坐标系下的高精度摄像机矩阵P。为任意角度的多个视图建立可扩展的主框架,不需要图像矫正。基于图像坐标和自动计算出的视差范围建立三维网络,为每条边分配适当的权值,把多视图立体匹配问题转化为网络最大流问题。算法保证了高准确率和平滑性,实验结果表明此方法适用于三维模型重建。The algorithm is decomposed into a number of successive steps to obtain precise depth map. Feature extraction and matching, computing F matrix, guided complementary matching, bundle adjustment etc. are first applied on an image sequence. Then the precise camera matrix for each image is extracted. The algorithm builds extendable main frame on the images with arbitrary viewpoints, constructs 3D network based on image coordinates and automatically computed disparity range, and assigns cost value for each edge. The multiview stereo matching problem is finally translated into the maximum flow problem. The algorithm ensures to high precision and smoothness in 3D scene reconstruction, and it is well performed in the experiment.

关 键 词：SIFT特征点 基础矩阵 摄像机矩阵 可扩展主框架 网络最大流 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]