Estimation of radial distortion using local spectra of planar textures

Abstract

A novel self-calibration method for estimation of radial lens distortion is proposed. It requires only a single image of a textured plane that may have arbitrary orientation with respect to the camera. A frequency-based approach is used to estimate the perspective and non-linear lens distortions that planar textures are subject to when projected to a camera image plane. The texture is only required to be homogeneous and may exhibit a high amount of stochastic content. For this purpose, we derive the relationship between the local spatial frequencies of the texture and those of the image. In a joint optimization, both the rotation matrix and the radial distortion are subsequently estimated. Results show that with appropriate textures, a mean reprojection error of 9.76 · 10−5 relative to the picture width is achieved. In addition, the method is robust to image corruption by noise.