Camera calibration and pose determination are two fundamental problems in computer vision. The thesis aims to explore new algorithms for the camera calibration and pose estimation using weighted least squares technique. The main points are summarized as follows: 1. A novel weighted linear method for the camera pose estimation is presented. The key idea of this method is to replace the algebraic error in the classic linear method with the weighted algebraic error to adequately approximate the geometric error. The method provides a linear solution whose accuracy is close to the accuracy of ML estimation. Based on the DLT and EPnP, the weighted DLT (WDLT) and weighted EPnP (WEPnP) are proposed. Experiments with simulative data and real images show that the WDLT and WEPnP remarkably outperform the DLT and EPnP and in the case of small depth ratio, both of them also outperform the Lu’s nonlinear algorithm. 2. A high-precision weighted linear 1D calibration algorithm is proposed. Since the accuracy of 1D calibration depends on the estimation of depth, based on the relationship between errors of depth and image distance, the distance between image points is used as the weight for each linear constraint in the classical 1D calibration, and the 1D calibration is transformed into a weighted linear least squares problem. A large number of experiments show that, with a comparable computational complexity, the calibration accuracy of our weighted algorithm is substantially improved.
修改评论