利用密度聚类估计基础矩阵的核线校正方法

The epipolar rectification method of base fundamental matrix estimated by density clustering

针对最小平方中值法(LMeds)和随机采样一致性(RANSAC)算法在立体像对核线校正过程中,基础矩阵容易受到错误定位和误匹配的影响,导致估算精度降低的问题,该文提出采用聚类分析估计基础矩阵的方法。该方法首先对匹配特征点进行聚类分析,获得更加准确可靠的局内点后,采用RANSAC算法估算基础矩阵,通过确定核点坐标完成立体像对间的快速映射,利用三次卷积法进行核线重采样,完成核线校正。选取3组影像进行核线校正实验,将所提算法与LMeds和RANSAC算法估计基础矩阵的核线校正结果相比,实验结果表明:所提算法消除同名点间上下视差的效果更好,其上下视差绝对值均在0.05个像素范围内,具有较好的稳定性与可行性,能够为立体像对的三维重建提供良好的基础。

In view of least median of squares(LMeds)and random sample consensus(RANSAC)algorithms in the process of stereo images epipolar rectification,fundamental matrix is susceptible to fault localization and mismatching,leading to the reduction of estimation accuracy.A method of estimating fundamental matrix by cluster analysis was proposed in this paper.The method firstly clustering analyzed the matching feature points to obtain more accurate and reliable interior points and then the RANSAC algorithm was used to estimate the fundamental matrix and the fast mapping between stereo images was completed by determining the epipole coordinates.Cubic convolution was used to carry out the epipolar resampling to complete the epipolar rectification.Three groups of images were selected for the epipolar rectification experiments and compared with the results of epipolar rectification which used LMeds and RANSAC algorithms to estimate fundamental matrix.Experimental results showed that the proposed algorithm eliminated the effect of parallax between the homonymy points,and the absolute value of the vertical parallax was within 0.05 pixels.This method had good stability and feasibility,and provided a good foundation for three-dimensional reconstruction.