Accurate camera calibration and feature based 3-D reconstruction from monocular image sequences

In this thesis, computational methods are developed for measuring three-dimensional structure from image sequences. The measurement process contains several stages, in which the intensity information obtained from a moving video camera is transformed into three-dimensional spatial coordinates. The proposed approach utilizes either line or circular features, which are automatically observed from each camera position. The two-dimensional data gathered from a sequence of digital images is then integrated into a three-dimensional model. This process is divided into three major computational issues: data acquisition, geometric camera calibration, and 3-D structure estimation. The purpose of data acquisition is to accurately locate the features from individual images. This task is performed by first determining the intensity boundary of each feature with subpixel precision, and then fitting a geometric model of the expected feature type into the boundary curve. The resulting parameters fully describe the two-dimensional location of the feature with respect to the image coordinate system. The feature coordinates obtained can be used as input data both in camera calibration and 3-D structure estimation. Geometric camera calibration is required for correcting the spatial errors in the images. Due to various error sources video cameras do not typically produce a perfect perspective projection. The feature coordinates determined are therefore systematically distorted. In order to correct the distortion, both a comprehensive camera model and a procedure for computing the model parameters are required. The calibration procedure proposed in this thesis utilizes circular features in the computation of the camera parameters. A new method for correcting the image coordinates is also presented. Estimation of the 3-D scene structure from image sequences requires the camera position and orientation to be known for each image. Thus, camera motion estimation is closely related to the 3D structure estimation, and generally, these two tasks must be performed in parallel causing the estimation problem to be nonlinear. However, if the motion is purely translational, or the rotation component is known in advance, the motion estimation process can be separated from 3-D structure estimation. As a consequence, linear techniques for accurately computing both camera motion and 3D coordinates of the features can be used. A major advantage of using an image sequence based measurement technique is that the correspondence problem of traditional stereo vision is mainly avoided. The image sequence can be captured with short inter-frame steps causing the disparity between successive images to be so small that the correspondences can be easily determined with a simple tracking technique. Furthermore, if the motion is translational, the shapes of the features are only slightly deformed during the sequence.