Investigation of the Effects of False Matches and Distribution of the Matched Keypoints on The PnP Algorithm

Perspective-n-Point (PnP) problem is the estimation of the pose (location and orientation) of a calibrated camera from 3D-2D point correspondences, that is, three-dimensional coordinates of objects in a world coordinate system and corresponding pixels in two-dimensional images. PnP algorithms are used in computer vision, augmented reality, robotics, photogrammetry etc. Distribution of the points in the image and the accuracy of the 3D information are important on the accuracy of the estimated pose of the camera. In this paper, keypoints with previously known coordinates and keypoints detected in images taken from a drone are matched. Then, the robustness of PnP algorithms is investigated by adding mismatched keypoints into the true matches. In addition, the effect of the distribution of the points in the image is investigated. The entire study was carried out using high-resolution orthophotos. Direct Linear Transformation (DLT), Efficient PnP (EPnP), LHM and Robust PnP (RPnP) were tested as pose estimation algorithms. As a result, it is observed that the RPnP and LHM algorithms performed better than the other pose estimation algorithms when mismatched keypoints are added. It is also observed that RPnP and LHM give accurate results when the distance error (between the true match and the false match) is increased. RPnP has an advantage over LHM in terms of computational cost. In the case of homogeneous distribution of keypoints, it is observed that PnP algorithms estimate more accurate position and orientation than in the case of nonhomogeneous distribution