Calibrating Depth Sensors with a Genetic Algorithm

Abstract

In this report, we deal with the optimization of the transformation estimate between the coordinate systems of depth sensors, \ie sensors that produce 3D measurements. For that, we present a novel method using a genetic algorithm to refine the six degrees of freedom (6 DoF) transformation via three rotational and three translational offsets. First, we demonstrate the necessity for an accurate depth sensor calibration using a depth error model of stereo cameras. The fusion of stereo disparity assumes a Gaussian disparity error distribution, which we examine with different stereo matching algorithms on the widely-used KITTI visual odometry dataset. Our analysis shows that the existing calibration is not adequate for accurate disparity fusion. As a consequence, we employ our genetic algorithm on this particular dataset, which results in a greatly improved calibration between the mounted stereo camera and the Lidar. Thus, stereo disparity estimates show improved results in quantitative evaluations.