Refine Camera Calibration with Global Geometry Constraints

Camera calibration is an important task in the computer vision community. Traditional calibration methods calculate the internal and external parameters of the camera by capturing images of the calibration board in different poses. However, the noise in the calibration images significantly affects the calibration results. Most conventional methods assume that the 2D detection of vision feature is accurate, which conflicts with the fact that image detection always contains noise. The reliance on vision features in the traditional calibration scheme could reduce the robustness of camera calibration due to image noise. This can lead to a situation where the reprojection error is small but the calibration results are not accurate. This paper proposes a novel camera calibration framework based on geometry constraints derived from a robotic arm. By combining data filtering with global geometry constraints during the calibration procedure, the accuracy of camera calibration is improved, and the impact of instability during image acquisition is reduced. Finally, extensive experiments show that the proposed mechanism can achieve better calibration results than traditional methods.