Enhanced calibration of camera setups for high-performance visual odometry

Abstract

Listen

Translate article

For robots and autonomous system that rely on visual data for operating in the real world, camera calibration is an indispensable step as it relates image information to the geometric structure of the 3D world. Although it is convenient to consider a several decades old problem as something that is swiftly solvable with a dedicated toolbox, we should still push calibration methods to their practical limits in order to gain valuable insights, and especially when robots are operating in circumstances that concern human safety. In this paper we propose a camera setup calibration procedure with emphasis on visual odometry accuracy. We focus on target-based calibration and two popular datasets are used for evaluating visual odometry and SLAM algorithms, namely the EuRoC and KITTI datasets. Our procedure consists of: (i) introducing a novel highly accurate corner detection algorithm robust to challenging illumination conditions, (ii) investigating different lens distortion models, (iii) incorporating static and dynamic board deformation models, (iv) ex-post analysis of reprojection error sensitivity and calibration parameter uncertainty, and (v) grid search method based on odometry accuracy when board poses do not constrain calibration parameters well enough. The whole process significantly reduced the reprojection error when calibrating the camera setups of the EuRoC and KITTI datasets. We tested four different odometries, namely SOFT, ORB-SLAM2, VINS-FUSION, and VISO2—all four showed higher accuracy with the proposed calibration parameters. Moreover, with the proposed calibration method our SOFT2 scored 0.53% in translation and 0.0009 deg/m in rotation error rendering it currently the highest ranking algorithm on the KITTI scoreboard.