Skip to Main Content
The extrinsic calibration of a stereo camera system is a procedure that estimates the position and orientation of a stereo camera system relative to a calibration object. In this paper, a calibration object specified by a CAD model of known shape and arbitrary texture is used. Based on feature points located in the stereo image and using triangulation, 3D object points in the camera coordinate system are calculated. This point cloud is matched to the surface of the CAD model to estimate the position and the rotation of the stereo camera system relative to the object coordinate system. In order to improve the accuracy of known techniques for the calibration, two steps are proposed in this paper. First, a systematic error in the standard method of subpel feature point localization is eliminated by replacing the parabolic interpolation by a Gaussian regression. Second, the estimated accuracies of the 3D points are incorporated by propagating the uncertainties of the detected feature points in the images. To calculate the global minimum of the cost function, an evolutionary optimizer is combined with a two-stage strategy, refining a fast converging approximate result. The developed procedure reduces the error variance of observed points by a factor of about 80.