期刊名称：ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
印刷版ISSN：2194-9042
电子版ISSN：2194-9050
出版年度：2008
卷号：XXXVII Part B5
页码：21-26
出版社：Copernicus Publications
摘要：Estimation of camera geometry represents an essential task in photogrammetry and computer vision. Various algorithms for recover- ing camera parameters have been reported and reviewed in literature, relying on different camera models, algorithms and a priori ob- ject information. Simple 2D chess-board patterns, serving as test-fields for camera calibration, allow developing interesting automa- ted approaches based on feature extraction tools. Several such 'calibration toolboxes' are available on the Internet, requiring varying degrees of human interaction. The present contribution extends our implemented fully automatic algorithm for the exclusive purpose of camera calibration. The approach relies on image sets depicting chess-board patterns, on the sole assumption that these consist of alternating light and dark squares. Among points extracted via a sub-pixel Harris operator, the valid chess-board corner points are automatically identified and sorted in chess-board rows and columns by exploiting differences in brightness on either side of a valid line segment. All sorted nodes on each image are related to object nodes in systems possibly differing in rotation and translation (this is irrelevant for camera calibration). Using initial values for all unknown parameters estimated from the vanishing points of the two main chess-board directions, an iterative bundle adjustment recovers all camera geometry parameters (including image aspect ratio and skewness as well as lens distortions). Only points belonging to intersecting image lines are initially accepted as valid nodes; yet, after a first bundle solution, back-projection allows to identify and introduce into the adjustment all detected nodes. Results for data- sets from different cameras available on the Web and comparison with other accessible algorithms indicate that this fully automatic approach performs very well, at least with images typically acquired for calibration purposes (substantial image portions occupied by the chess-board pattern, no excessive irrelevant image detail).
