Graphics enhanced version of this site
Accurate Camera Calibration from Planar Patterns
Heads: Takeo Kanade and Ankur Datta
Contact: Ankur Datta (adatta@andrew.cmu.edu)
Mailing address:
Carnegie Mellon University
Robotics Institute
5000 Forbes Avenue
Pittsburgh, PA 15213
Associated center: VASC
Associated lab/group: People Image Analysis Consortium
We have developed a novel camera calibration algorithm that utilizes a single framework for square, circle, and ring planar calibration patterns. The algorithm extends the previous work on camera calibration, which involved a one-step solution for the calibration parameters to a multi-step refinement. We utilize the initial calibration parameters obtained from one-step algorithm to perform undistortion and unprojection of the calibration pattern to a canonical fronto-parallel plane, which is then used to perform corner detection and solve for calibration parameters. This process is repeated again with the updated calibration parameters till convergence. Projecting the calibration grid to the canonical plane dramatically increases the accuracy of corner detection and consequently of the camera calibration estimate. We have conducted an extensive set of experiments with real and synthetic images and the pixel reprojection errors obtained by our method are about 50% lower as compared to the OpenCV Camera Calibration Toolbox. As a possible side-benefit, increases in accuracy of intrinsic camera parameters directly leads to increases in accuracy of stereo camera calibration as well.
| Improved Results: Iterative Approach + Pattern
| |
| Visual Hull Reconstruction Result: ant | |
| Ground Truth (video)
|
OpenCV (video) |
| Ring (video)
| Circle (video) |
| Visual Hull Reconstruction Result: seaweed | |
| Ground Truth (video) |
OpenCV (video) |
| Ring (video) |
Circle (video) |