Direct 3D metric reconstruction from two views using differential evolution

To obtain a 3D metric reconstruction from two images taken with a same camera without previous calibration, it is necessary to estimate the intrinsic camera parameters and the orientation and position of the two views with respect to the camera. At the present time, there are several algorithms to estimate camera parameters from two views, all of them are based on the epipolar geometry and the estimation of the fundamental matrix. However, it is well known there are some configurations where the fundamental matrix can not be estimated, called critical configurations. In this article we present a novel method to retrieve directly the camera parameters, and orientation and position parameters for two views, from points taken over the two images, using the Differential Evolution (DE) algorithm. This method exploits the reprojection error as the cost function for DE, instead of computing the fundamental matrix. Experimental results show our method recovers 3D points, intrinsic, and orientation and position parameters on non-critical configurations and in the critical configuration of pure translation.We used simulated and real images to prove its effectiveness and robustness.

I. Introduction

One of the main problems in Computer Vision, which has been deeply studied, is the three-dimensional metric reconstruction from two uncalibrated views [1], [2], [3], [4], [5], [6], [7], [8], [9].