This paper proposes a robust calibration algorithm integrating multi-factor and normalization optimization, targeting the localization deviation issue in binocular vision calibration caused by defocus blur, perspective distortion, and noise interference of anchor points on the imaging plane due to the convex lens principle. By quantifying the errors from defocus blur, degree of distortion, and signal-to-noise ratio, assigning different weights to these multi-factor errors, and implementing a local normalization strategy, an error weight formula is defined to eliminate anchor points with large errors and suppress the interference of high-deviation anchor points on the overall calibration results. Experiments show that compared with Zhang's calibration method, this approach reduces the average reprojection error by 30% in near, medium, and far viewing fields, and decreases the errors of the rotation matrix and translation matrix by 4.5% and 3.4%, respectively. Moreover, it achieves good localization accuracy at different tilt angles and significantly enhances robustness in complex scenarios. This research provides a high-precision calibration solution for binocular stereo vision systems in fields such as SLAM, robot localization, autonomous driving, and industrial inspection. Future work will extend this method to larger-scale complex environments to verify its practicality.
Yipeng Zhao (Fri,) studied this question.