Camera Modelling and Calibration for Machine Vision Applications
Licentiate thesis, 2008
This thesis describes camera based methods for determining positions and orientations of objects in 3D-space. These methods are useful in several applications where camera images are used for acquiring information about the surroundings, and the targeted application here is guiding robots using a camera mounted on the robot hand. In general there can be one or more cameras involved, but in any case there is need for a calibrated camera model and efficient algorithms in order to perform accurate pose calculations. There is further need for at least two images of the scene in order to obtain position and orientation of objects in the scene.
The thesis motivates the need for and proposes a new camera model together with calibration routines and applications.
The new generic camera model takes into consideration normal camera distortions like radial distortion, but also distortions due to varying entrance pupil and decentring distortion. Further, the proposed new camera model can handle fisheye cameras with wide angle lenses as opposed to conventional camera models. It is also suitable for situations where the focus and zoom of the camera can vary.
The camera calibration makes use of references in the surroundings. The camera takes a number of images of the references from different angles and distances. Then the calibration calculates camera parameters from these images, using optimization routines. The references' positions can either be known in advance or unknown. In the latter case the references are also calculated in the calibration. Difficulties arise e.g. when the same image can be obtained with different parameter values, and strategies to overcome this is presented. For the optimisation to converge it needs reliable initial values, and methods for estimation of these are based on the shape, size, and localisation of the references in the images. A method for determining image centrepoints of the references is also described. These are needed for higher accuracy of the vision system.
The text assumes that 2D-images are acquired by suitable means and that image processing techniques are known to the reader.