Research Of Global Motion Estimation And Camera Calibration Algorithm

As the continuous development of multimedia technology in recent years, video compression has become a hotspot in relevant fields. Motion estimation/compensation technique is the key technology in the framework of video compression, which has the greatest contribution to the compression ratio and quality.Stereo vision, which is an important branch in the emerging disciplines of computer-vision, is in rapid development. Camera calibration with high accuracy is the premise of reliable 3D information reconstruction and modeling. Consequently, research on camera calibration and its application in 3D information extraction is of theoretical and practical importance.The main content of this paper consists of two parts: global motion estimation and camera calibration.This paper details the parametric model of the motion vector field and analyzes the direct/indirect method for global motion estimation. This paper also points out the contradiction between estimation precision and complexity in the GME problem.Aimed at the problems in the GME procedure, a pre-processing and post processing approach for fast global motion estimation is proposed. Six-parameter affine motion model with both "minimal motion vector residual" and "minimal luma residual" as optimization criteria is used. Based on the "minimal MV residual" criteria, principle and methodology of feature point selection in pre-processing is proposed, which plays an important role in image registration. Outliers are removed efficiently in the form of "seed-proliferation". Based on the "minimal luma residual" criteria, a fast parameter refinement method is proposed, in which only a small subset (down to 2%) of the original image pixels involve in post-processing procedure. Experiments show that even in the case of image-blue due to defocusing or fast motion, the proposed refinement method performs well to modify the parameters with bad initial values.As regarding the camera calibration, the linear and nonlinear camera models are described in details. Several classic methods is described in details, e.g. the direct linear transform (DLT) method based on linear camera model and the Zhang Z.Y. 's method based on Two-step routines. In this paper, a method for stereo calibration is proposed.