In this thesis, we propose a new multi-viewpoint camera calibration method for the voxel-based scene reconstruction. The proposed method estimates the extrinsic parameters of a camera assuming that the intrinsic parameters are known and the minimal three-dimensional information is given.
The proposed method is composed of three steps. First, We use the initial two images that contain the known three-dimensional information to estimate the initial extrinsic parameters. Second, we estimate the extrinsic parameters of other viewpoints sequentially with the computed three-dimensional points in the previous step. Finally we use the estimated epipolar geometry to increase the matching points, and gradually reduce the calibration error with increased matching points.
The optimization process adopted for the whole estimation is composed of a process to estimate a linear solution and two optimization methods to refine the linear solution. The given three-dimensional information of a minimum number of six scene points analytically determines an initial estimate of extrinsic parameters.
Subsequently, the combination of two optimization methods reduces the reprojection error starting from the initial estimates. We further reduce the accumulated error by dividing the entire image sequence into two parts.
The calibration method is validated with various kinds of input images captured from synthetic data, objects in indoor scene and outdoor environments. The experimental results demonstrate that the proposed method can be practically applied for outdoor scene reconstruction without using any calibration patterns.
