Camera Calibration by Corner Detection

¤   Problem Statement

The goal of this project is, given a prewritten and relatively efficient program for calibrating a digital camera based on very little material, to try to improve its accuracy without necessitating the use of more input.

¤   Background

Camera calibration is the name for estimating several useful characteristics of a digital camera based on some type of input from the camera itself. The procedure finds the camera's position in 3-space relative to a point fixed with respect to one or more of the images; the camera's internal error, e.g. from lens distortion; and the homographies (3-D rotations and translations) that geometrically connect the different image planes taken by the camera.
See a sample image of the calibration aid; the papers I have read on current methods and the need for improvement.

¤   Purpose

The Virtual Reality Laboratory at NRL would like to be able to easily calibrate a camera with only a few (eight at the moment) images of a two-dimensional calibration object. Their current calibration program is not nearly as accurate as they would like it to be.

¤   Scope

I will be working with the researcher who tutored me over the summer at the Naval Research Lab, Dr. Mark Livingston (Virtual Reality Laboratory). I will first rewrite the calibration algorithm I will be modifying, originally written by Zhengyou Zhang of Microsoft Research, to be easily customizable; then I will create and attempt various improvements to his method to increase its accuracy.

¤   Procedure

After rewriting Zhang's procedure to my satisfaction, I will, under Mark's guidance, test out ideas for improving the performance of the calibrator, probably based on modifying the input data (the detected locations of corners in images).

See a PDF article about the Harris method (pg. 4 - 7), a popular method for corner detection that involves the use of the image gradient. I will most probably be following Harris' lead in using the gradient to adjust the detected positions of corners before feeding them to the calibration procedure.

¤   Expected Results