Design and application of an automated system for camera photogrammetric calibration

Doctoral Thesis


Permanent link to this Item
Journal Title
Link to Journal
Journal ISSN
Volume Title

University of Cape Town

This work presents the development of a novel Automatic Photogrammetric Camera Calibration System (APCCS) that is capable of calibrating cameras, regardless of their Field of View (FOV), resolution and sensitivity spectrum. Such calibrated cameras can, despite lens distortion, accurately determine vectors in a desired reference frame for any image coordinate, and map points in the reference frame to their corresponding image coordinates. The proposed system is based on a robotic arm which presents an interchangeable light source to the camera in a sequence of known discrete poses. A computer captures the camera's image for each robot pose and locates the light source centre in the image for each point in the sequence. Careful selection of the robot poses allows cost functions dependant on the captured poses and light source centres to be formulated for each of the desired calibration parameters. These parameters are the Brown model parameters to convert from the distorted to the undistorted image (and vice versa), the focal length, and the camera's pose. The pose is split into the camera pose relative to its mount and the mount's pose relative to the reference frame to aid subsequent camera replacement. The parameters that minimise each cost function are deter- mined via a combination of coarse global and fine local optimisation techniques: genetic algorithms and the Leapfrog algorithm, respectively. The real world applicability of the APCCS is assessed by photogrammetrically stitching cameras of differing resolutions, FOVs and spectra into a single multi- spectral panorama. The quality of these panoramas are deemed acceptable after both subjective and quantitative analyses. The quantitative analysis compares the stitched position of matched image feature pairs found with the Shape Invariant Feature Tracker (SIFT) and Speeded Up Robust Features (SURF) algorithms and shows the stitching to be accurate to within 0.3°. The noise sensitivity of the APCCS is assessed via the generation of synthetic light source centres and robot poses. The data is realistically created for a hy- pothetical camera pair via the corruption of ideal data using seven noise sources emulating the robot movement, camera mounting and image processing errors. The calibration and resulting stitching accuracies are shown to be largely independent of the noise magnitudes in the operational ranges tested. The APCCS is thus found to be robust to noise. The APCCS is shown to meet all its requirements by determining a novel combination of calibration parameters for cameras regardless of their properties in a noise resilient manner.