基于软指派算法的相机位姿估计研究

Camera Pose Estimation Based on Softassign Algorithm

针对未知特征点匹配关系下的单目相机位姿估计问题,提出了一种基于软指派算法的相机位姿估计算法。结合了正交迭代算法和软指派算法,以三维(3D)和二维(2D)特征点的物方空间共线性误差作为确定特征点之间匹配关系的指派依据和计算相机位姿的目标函数值,通过迭代方式确定特征点的匹配关系和相机的位姿。其不仅能处理3D/2D特征点一一对应的情况,而且能够处理同时存在遮挡的3D特征点和错误的2D特征点的情况。合成图像的实验结果表明:在含有图像噪声、遮挡的3D特征点和错误的2D特征点情况下,该算法的成功率>82%,真实图像实验对算法的性能进行了验证。

Aiming at the problem of monocular camera pose estimation under the matching relations of unknown feature points, a new camera pose estimation algorithm is proposed based on the Softassign algorithm. Based on the advantages of orthogonal iteration and Softassign algorithm, the collinear error of the object space in three- dimensional （3D） feature points and two dimensional （2D） feature points is used as the basis for determining the matching relationship between feature points and the objective function value for calculating the camera pose. A iteratively way determines the matching relationship of feature points and the pose of the camera. It can handle the case where the 3D/2D feature points are in one-to-one correspondence, and the case which occlude 3D feature points and erroneous 2D features. The results of the synthetic image experiments show that the success rate of this algorithm is over 82% in one case with image noise, occlusion and clutter. The experiment of real image can verify the property of algorithm.