基于未标定图像序列的稠密重建算法

A Dense Reconstruction Algorithm Based on Uncalibrated Images

提出一种基于未标定图像序列的稠密重建算法。算法首先利用SIFT特征点对应估计基本矩阵,同时计算图像匹配视差参考值。然后利用由基本矩阵计算得到的对极关系对图像进行平行极线修正,结合视差参考值对经过rank变换后的图像进行稠密匹配。最后自标定照相机,优化对应点的2D-3D投影关系,重建场景三维结构。实验结果表明,本算法能够对有效图像(考虑宽基线引起的大视差情况)80%以上的像素实现准确的匹配,重建出稠密的三维空间点云。

In this paper,a dense reconstruction algorithm based on uncalibrated image sequences is proposed.SIFT feature points are detected and matched from original images.Fundamental matrix is robustly estimated using RANSAC algorithm. To improve the accuracy of dense match,original images are rectified based on epipolar geometry which calculated according fundamental matrix,and transformed to'rank images'.Finally,camera is self calibrated and 3D scenery is reconstructed by 2D-3D reprojection.The experimental results proved that the dense point correspondences were correct for 80%image pixels and dense 3D points cloud could be reconstructed by the proposed algorithm.