摘要
特征匹配是从图像恢复三维模型的关键步骤之一.为有效地提高三维重建的质量,提出一种面向三维重建的增强运动一致性与引导扩散特征匹配算法.首先在基于网格的运动统计算法基础上,通过增加阈值β,提出一种增强运动一致性概念,增强真假匹配点的判断条件,避免高相似特征点的误匹配,提高了初始匹配点的正确率;然后结合RANSAC算法进行特征点匹配优化,过滤掉异常值,进一步提高特征点匹配的准确性;最后将引导匹配和运动一致性相结合,提出一种引导扩散概念,减少了集中分布在图像局部的可能性,进而提高特征点匹配数量和三维模型的稳定性.在公开的三维重建数据集的618对图像上的实验结果表明,该算法在特征匹配和三维重建上能够实现更好的性能;在小于1°误差阈值的位姿估计成功率上,该算法比基于SIFT的比率测试算法和GMS算法分别平均提高了22.58%和12.90%,尤其在重复纹理图像对上分别提高了46.15%和30.77%.
Feature matching is one of the key steps to restore a 3 D model from an image. To effectively improve the quality of 3 D reconstruction, an enhanced motion consistency and guided diffusion feature matching algorithm for 3 D reconstruction is presented. Firstly, based on the grid-based motion statistics algorithm, an enhanced motion consistency concept is proposed by adding a threshold β, which enhances the judgment condition of true and false matching points, avoids the false matching of highly similar features,and improves the initial matching points accuracy. Then, the RANSAC algorithm is used for feature point matching optimization to filter out outliers and further improve the feature point matching accuracy. Finally,a guided diffusion concept that combines guided matching and motion consistency is proposed, which reduces the possibility of concentrated distribution in the part of the image, thereby improving the feature points matching number and the 3 D model stability. Experiments on 618 pairs of images in the public 3 D reconstruction datasets demonstrate that this algorithm can achieve better performance in feature matching and 3 D reconstruction. For the success percentage of pose estimation less than 1° error threshold, the proposed algorithm is 22.58% and 12.90% higher than the SIFT-based ratio test algorithm and the GMS algorithm, respectively. In particular, it is 46.15% and 30.77% higher on repeated texture image pairs.
作者
蔡振姣
张素兰
李晓明
张继福
胡立华
杨海峰
Cai Zhenjiao;Zhang Sulan;Li Xiaoming;Zhang Jifu;Hu Lihua;Yang Haifeng(College of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024)
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2022年第2期273-282,共10页
Journal of Computer-Aided Design & Computer Graphics
基金
国家自然科学基金(U1731126)
大数据分析与并行计算山西省科技创新重点团队项目(201805D131007).
