Iterative Linear Solution of the Perspective n-Point Problem Using Unbiased Statistics

In this paper, we cope with the problem of pose estimation of a calibrated camera from n 3D-to-2D point correspondences (PnP). In the state of the art, there exists an efficient solution to this problem (EPnP) with linear computational complexity; nevertheless, the solution is obtained as a linear combination of four eigenvectors of a matrix and the linear coefficients are not always straightforward determined. The EPnP algorithm uses a Minimum Least Squares (MLS) solution that is biased for low signal-to-noise measurements.We propose an alternative implementation to the EPnP algorithm based on Kanatani’s unbiased statistical estimators. Kanatani’s approach has shown to perform better than MLS for noisy data. Our method takes into account the statistical distribution of the data and the solution is straightly computed from the null-space of a modified linear system, preserving the linear computational complexity as it is shown in the experiments.