Stability analysis of intrinsic camera calibration using probability distributions

The currently most widely used calibration method in geometric computer vision is based on Zhang's method. This approach is easy to implement and delivers detailed result evaluations. One drawback, however, might be that the unexperienced user may obtain unstable and unrepeatable calibration results despite small re-projection errors. Therefore, the data must be interpreted with caution and there is definitely a need for further consideration beside re-projections to be taken for better estimations. So far, statistical inference has been used for presenting precision measures only. In our work, we extend the statistical approach quantitatively using large bundles of images and get calibrations from randomly chosen subsets. Then a Maximum Likelihood Estimation of intrinsic parameters is implemented and the statistical behaviour is analyzed. In addition, the recovered expected values of parameters are utilized as ground truth to scrutinize the single influencing factors of the imaging configuration. According to the results we found out that the image block plays a significant role for camera calibration with regard to the orientation of the imaging configuration. Including also suitable manufactured calibration boards as well as other operational issues, finally a well-designed image block provides a repeatable and reliable calibration.