Robust fourier-based checkerboard corner detection for camera calibration

Abstract

Precise localization of reference markers is crucial for the accuracy of target-based camera calibration. State-of-the art detectors, however, are sensitive to optical blur corrupting the image in many practical calibration scenarios. We propose a novel method for the sub-pixel refinement stage of common checkerboard target detectors. It uses the symmetry of checkerboard crossings and exploits the periodicity in the angular frequency domain when the origin of a polar coordinate system is centered at the crossing. The detector estimates the crossing center’s sub-pixel position by minimizing spurious frequency components that occur increasingly at ever larger distances from the crossing center. An average localization error of 0.08 px is achieved in noisy and artificially blurred synthetic images, surpassing the state of the art by 65 %. In addition, we evaluated the detector in real-world camera calibration using a public data set, achieving an reprojection error of 0.11 px compared to 0.27 px for the state of the art.