| For common situations in autonomous driving scenarios, we propose a planar relative pose estimation method based on a single SIFT feature, which further reduces the motion to two degrees of freedom with only rotation angle and translation angle by limiting the camera movement to a plane motion of three degrees of freedom. At the same time, considering that the homography information of the ground plane contains the information of the camera motion and the plane normal vector between the two images, the corresponding relationship between the ground points between the two images is described, and the camera motion can be restored by estimating the homography matrix. In order to reduce the number of point pairs used, we further introduce SIFT feature points, a pair of SIFT feature points includes the image coordinates of the corresponding points in the two images, as well as their feature rotation and feature scale, which greatly expands the information contained in a single point pair. Finally, the estimation of the relative pose of the camera is completed using only one SIFT feature, and finally the random sampling consensus algorithm is used to optimize the results to obtain the final result. We compare errors with other comparison methods in simulation experiments and real experiments, which proves that the proposed method is relatively reliable, effective and efficient. |
