Camera Modelling and Calibration with Machine Vision Applications

Doktorsavhandling, 2010

Camera modelling and calibration are important parts of machine vision. They can be used for calculating geometric information from images. A camera model is a mathematical projection between a 3D object space and a 2D image. The camera calibration is a mathematical procedure calculating parameters of the camera model, usually based on several images of reference points. These fundamental parts of machine vision are improved in this thesis. One large part is the development of a generic camera model, GCM, that is accurate, computationally efficient and can be used for both conventional, fisheye and even catadioptric cameras. Different models were used in the past for conventional and omnidirectional cameras and this is a well-known problem, the solution of which is described in this thesis. The accuracy of camera models is improved by introducing new ways of compensating for different distortions, such as radial distortion, varying entrance pupil point and decentring distortion. Calibration is improved by introducing new means of calculating start estimates of camera parameters, from analysing shapes, sizes and positions of the reference points in the images. These start estimates are needed in order to make the calibration converge. Methods for calculating better reference centre points than the centres of gravity are developed in order to increase the accuracy further. Non-trivial null spaces that occur during calibration are identified. Awareness of these improve the calibration. Calibrations with different camera models are implemented and tested for real cameras in order to compare their accuracy. Certain models are better for certain situations, but the overall performance and properties are favourable for the GCM. A stereo vision welding robot system is developed, using the new model. It determines the geometry of a 3D weld joint, so that a robot can follow it. The same system is implemented in a virtual environment using a simulation software. Such simulation is important since it makes it possible to develop robot vision systems off-line.