Student: Shih, Sheng-Wen Advisor: Lin, Wei-Song; Hung, Yi-Ping

This work provides thorough accuracy analysis for calibration of reconfigurable binocular heads. Based on the results of our analysis, we have developed a four-stage calibration method consisting of motorized lens calibration, kinematic calibration, head/eye calibration and kinematic parameter refinement. Extensive computer simulations have been performed to verify our theoretical analysis. To demonstrate the accuracy of the four-stage calibration method, real experiments were conducted by using an eight-joint binocular head built with off-the-shelf components, which comprised two joints for horizontal translation, two joints for pan and tilt, two joints for the left and right camera verging control and two joints for the focus control. The results of the real experiments have shown that the four-stage calibration method can achieve a very high accuracy which will cause only one pixel prediction error and 0.2 pixel epipolar error, on the average, even when all the eight joints are moved simultaneously. So far this is the most accurate results attainable in the literature on calibrating a fully reconfigurable binocular vision system, and this high accuracy is extremely useful for alleviating the difficulties encountered in many 3-D vision applications such as automatic reconstruction of 3-D objects and environment using an active stereo vision system. In this thesis, we first present a camera calibration method and two theoretical analyses on camera calibration error. The first error analysis is to investigate the effects of neglecting lens distortion in camera calibration. The second analysis is to derive the estimation accuracy of the physical camera parameters with respect to four different types of camera calibration problems. Based on this analysis, we have proposed a camera model which is suitable for calibrating motorized-focus lens, and have used it in a new method for calibrating a motorized-focus lens, which provides very accurate estimates of camera parameters for variable focus settings. For the second stage, we have developed two closed-form solutions for calibrating the kinematic parameters of binocular heads, where one uses the pose measurements and the other uses the point measurements. These two closed-form solutions can also be used to calibrate any serial robot (either a robot arm or a robot head) with or without multiple end-effectors. Theoretical error analyses for these two closed-form solutions have been given to provide a guideline for selecting an appropriate kinematic calibration method and for reducing the calibration error. For the third stage, a simple method is used to calibrate the head/eye relation by using a coordinate measurement machine. In the last stage, a recursive nonlinear least-square estimator is used to refine all the kinematic parameters (including the head/eye relation) to achieve high enough accuracy for 3-D reconstruction. The major contribution of this work is in the derivation of a series of accuracy analyses which provide better insights into the calibration problems concerning reconfigurable binocular vision systems. Another significant contribution is that, based on our error analysis, we have proposed a four-stage calibration method which can easily achieve very accurate calibration information that is important for many 3-D vision applications (but is often obviated by most computer vision researchers due to its complexity).