基于共面直线迭代加权最小二乘的相机位姿估计

针对相机内参数已标定和未标定情况下的相机位姿求解,提出了基于共面直线迭代加权最小二乘的相机位姿估计算法。推导了关于相机焦距和位姿参数的线性方程,通过4条以上共面直线实现了相机位姿参数的线性解算;对参数线性解进行迭代加权最小二乘优化,得到更高精度的参数估计值和直线权值;最后,利用直线权值和欧式变换的保距性实现相机焦距的解算,得到了相机焦距和位姿参数的估计值。仿真实验表明:提出的算法在相机已标定,直线数为20,像点噪声方差为5pixel的情况下,角度误差小于0.2°,相对平移向量误差小于0.5%,耗时大约为1ms。真实数据实验表明,提出的算法可以获得与棋盘标定结果相近的精度。与现有算法相比,提出的算法抗噪性更好,精度更高,能够实现基于单幅图像的未标定相机的位姿估计。

Iteratively reweighted least squares method for camera pose estimation based on coplanar line correspondences

To estimate the camera poses for calibrated or uncalibrated cameras, a novel pose estimation algorithm was proposed based on coplanar line correspondences and iteratively reweighted least squares. Firstly, a linear equation for the focus length and pose of the camera was established. The pose parameter was solved by more than four coplanar line correspondences. Then, the iteratively reweighted least square method was applied to optimizing the parameter, and the higher accurate estimated parameter and line weight were obtained. Finally, the focus length and pose parameters of the camera were obtained by calculation of line weight and the invariant distance of Euclidean transformation. Experimental results with simulative data indicate that the precision of angle is better than 0.2°, the precision of relative position is better than 0.5%, and consuming time is about 1 ms, when the focus length is known, and number of lines and noise level are 20 and 5, respectively. Moreover, the experimental results with real data indicate that the precision of proposed algorithm is close to chessboard calibration. As compared with the existing algorithm, the proposed algorithm is more accurate, robust, and is capable of estimating the pose of uncalibrated cameras based on a single image.