BibTeX entry

@PHDTHESIS{200707David_Claus,
  AUTHOR={David Claus},
  TITLE={High Accuracy Metrology Using Low-Resolution Cameras},
  SCHOOL={Oxford University},
  MONTH=Jul,
  YEAR=2007,
  URL={http://www.bmva.org/theses/2007/2007-claus.pdf},
}

Abstract

Camera based localization can provide extremely accurate 3D pose information, even from consumer grade video lenses. Advances in lens distortion correction, pose computation and feature detection would permit low cost cameras to be used in many applications that currently require more expensive equipment. I show how: (1) careful modelling and (2) careful fitting of these models to data; provides increased camera accuracy from the same camera equipment with little or no additional computational overhead. The primary contribution towards camera modelling is a lens distortion model based on rational functions that can represent standard, fisheye and catadioptric lens systems. Three separate calibration methods are demonstrated, making this a useful technique that can be implemented in a wide range of applications. Evaluation of calibration precision indicates that the proposed model accurately represents real-world lens distortion and provides lower errors than other models in common use. Although sensitivity to image noise can be a problem with such flexible models, several techniques are presented here that yield robust calibration in the midst of image uncertainty. I demonstrate multiple view camera auto-calibration on fisheye lens sequences using point correspondences alone, without first requiring the removal of lens distortion. Fitting of the camera model is improved by including a non-linear optimization to tune the model parameters against a known error measure. Careful optimizer construction is shown to avoid local minima, converge in realtime and achieve very high levels of precision. Image feature detection error is transmitted through the entire calibration process, so a robust exemplar based learning scheme is proposed to accurately detect known fiducial markers. This efficient classification approach handles the challenges of changing scene conditions (lighting variation, motion blur, clutter) without the large increase in false detections that plague other detection algorithms.