球形全景影像位姿估计的改进EPnP算法

球形全景成像可以克服透视成像视场角的局限,实现场景全覆盖的三维重建和量测。本文在普通影像位姿估计的EPnP(efficient perspective-n-point)算法上进行了改进和扩展,提出了一种稳健快速的球形全景影像位姿估计算法。首先,构建球形全景影像的投影模型,将EPnP算法的平面透视成像模型扩展到球面成像模型;然后,采用基于全景球心、像点、控制点共线条件方程的改进EPnP算法求解控制点的球形全景像空间坐标;最后,利用Horn绝对定位算法直接解算全景影像位姿。与球形全景影像位姿估计的后方交会算法的对比试验结果表明,本文提出的方法无须迭代求解,更为稳健快速,即使控制点数目较少也能达到高精度,基于非严格共中心拼接的全景相机,重投影误差可控制在3.00像素左右。

Position and Pose Estimation of Spherical Panoramic Image with Improved EPnP Algorithm

Spherical panoramic imaging can overcome the limitations of perspective imaging's range of view and achieve three-dimensional reconstruction and measurement with full coverage.The EPnP(efficient perspective-n-point) algorithm fit for ordinary image's position and pose estimation is improved and extended by this paper to put forward a robust and fast method to estimate spherical panoramic image's position and pose. Firstly,we construct the projection model of spherical panoramic image to extend the perspective plane imaging model of EPnP algorithm to the spherical imaging model;then with the improved EPnP algorithm basing on the condition equation that panoramic sphere center,image point and control point are collinear, we obtain the control points' coordinates in spherical panoramic image space; finally, we solve the spherical panoramic image's position and pose directly with the Horn absolute positioning algorithm. Compared with the position and pose estimation method of spherical panoramic image with the resection algorithm, the experiment results show that the method proposed in this paper can achieve higher robustness, higher efficiency without going through an iterative solution, and it can also achieve high accuracy with a few control points.Based on the panoramic camera stitching with non-strict center, the reprojection error can be controlled around 3.00 pixel level.