基于单个SIFT特征的相对位姿估计方法

Relative pose estimation method based on single SIFT features

位姿估计是精密光测和自动驾驶的基本问题之一。针对自动驾驶等实际应用中,相机在平面上运动,相机位姿的自由度为3的情况,本文提出了基于单个SIFT特征的相机相对位姿估计方法。由于单目相机无法恢复平移尺度,因此相机运动的自由度减少为仅有旋转角和平移角的两自由度。通过观测地面,可以得到包含相机运动和平面法向量的地面单应信息,因此可以通过提取地面同名点估计单应矩阵来恢复相机运动。为了减少RANSAC迭代次数、提高算法效率,引入SIFT特征进行位姿估计。SIFT特征包括2幅图像中同名点图像坐标以及其特征旋转和特征尺度,可以扩充单个点对中包含的信息,有效减少求解单应矩阵所需点对数量。针对平面二自由度运动情况,本文使用单个SIFT特征点对完成单应矩阵的估计,并采用随机采样一致算法对结果进行优化,最终分解单应矩阵得到相对位姿估计结果。在仿真实验及真实实验中与2pt方法和5pt方法进行对比,证明了所提出的方法是有效的。

Position estimation is one of the fundamental problems in precision optical measurement and autonomous driving.For practical applications such as autonomous driving,where the camera moves on a plane and the degrees of freedom of the camera position are three,this paper proposes a camera relative pose estimation method based on a single SIFT feature.Since the monocular camera cannot recover the translation scale,the degrees of freedom of the camera motion are reduced to two degrees of freedom with only rotation and translation angles.By observing the ground,ground homography information containing camera motion and plane normal vectors can be obtained.Therefore,camera motion can be restored by extracting homonymous ground points to estimate the homography matrix.In order to reduce the number of RANSAC interactions and improve the efficiency of the algorithm,SIFT features are introduced to the pose estimation,which include the coordinates of the homonymous points in the two images,as well as their feature rotations and feature scales.So that the information contained in a single point pair can be expanded,and the number of point pairs required for solving the homography matrix can be reduced efficiently.For the case of planar two-degree-of-freedom motion,this paper uses a single SIFT feature point pair to complete the estimation of homography matrix,and then uses Random Sample Consensus algorithm to optimize the results,and finally decomposes the homography matrix to obtain the relative position estimation results.The proposed method is proved to be effective by comparing it with the 2pt method and 5pt method in simulation and real experiments.