Optimization methods to calibrate a stereo rig with increased accuracy for vehicular applications

In this paper, several methods are presented to fully calibrate a stereo vision system for far-range outdoor applications. Traditionally, intrinsic camera parameters are determined from a high number of feature points, while pose estimation is solved with a few (tens). Close-range setups are not well suited for far-range applications, while far-range setups can easily deviate from planar and small measurement errors may result in large errors in the marker positions. Our Maximum Likelihood (ML) formulation to the stereo pose problem takes such inaccuracies into consideration. Herein, our earlier results are extended by decoupling pose estimation into an inter-camera and a rig-to-world pose problem to avoid relying on a small number of features with all extrinsic parameters. A high number of point-matches are extracted from image pairs acquired during on-line operation, and are incorporated into the procedure for off-line inter-camera pose estimation via the essential matrix. Using real and synthetic data, it is shown how the far-range arrangement and the matches can be used to adjust the camera models, even after a sound intrinsic calibration and ML pose estimation. The proposed methods rely on point matches between views and, thus, are expected to produce better reference for the evaluation of autocalibration methods that typically begin with establishing such correspondences.