| Zoom-lens cameras have great potential in the applications of active vision. In such applications, the zoom, focus and aperture of a lens can be controlled to adapt to different lighting conditions or to obtain the desired field of view, depth of field, spatial resolution, or focused distance. Although a zoom-lens system is more flexible and useful than a fixed-parameter lens, it is not an easy job, in general, to calibrate a zoom-lens. The goal of this calibration is to determine the relationships between the lens settings (control parameters for the driving motors) and the different camera model parameters.; This dissertation develops a novel framework to produce an adjustable, perspective projection camera model of zoom-lens cameras across the lens control settings. The proposed framework is flexible enough to represent complex functions on continuous ranges of lens control space. This framework consists of a number of multi-layered feedforward neural networks (MLFNs) learning, concurrently, the perspective projection transformation of the camera across its control space.; Another important aspect of the calibration problem is lens distortion, which is often the main cause of deviation from the ideal pinhole camera geometry. In this work, lens distortion is calibrated and removed from the input images before any further processing, thus allowing the camera to be considered as an ideal pinhole camera. To calibrate lens distortion, we have developed new techniques based on the analysis of distorted images of straight lines. In particular, we derive new distortion measures that can be optimized using non-linear search techniques to find the best distortion parameters that straighten these lines. We also provide fast, closed-form solutions to estimate the distortion coefficients. Moreover, while almost all existing distortion calibration methods need user involvement in one form or another, we present a robust approach to lens distortion calibration based on the-least-median-of-squares (LMedS) estimator. Our approach is thus able to proceed in a fully-automatic manner while being less sensitive to erroneous input data such as image curves that can be mistakenly considered as projections of 3D linear segments.; The overall calibration framework was tested on real automated zoom-lens systems, operated across continuous ranges of focus and zoom with a root mean square error of half a pixel between the true and the measured positions of features in the image plane. Our approach demonstrates better performance than the current standard approach of Willson. Being part of an active vision system developed at the CVIP lab, the CardEye system, the calibrated zoom-lens cameras are used to build 3D Euclidean reconstruction of objects imaged by the system at the lens settings that provide best imaging conditions. |
