Research On Camera Pose Estimation Method Based On Scene 3D Surface Model

Recently,augmented reality,robotics navigation,simultaneous localization and mapping et al attract much attention of both academic and industrial community,in which image-based camera pose estimation is a crucial task.Camera pose is to infer the camera's spatial position and rotation based on the images obtained from the 3D scene.There has an excellent overview on some aspects of the image-based camera pose estimation and the related techniques are introduced.The method of recovering camera pose based on quaternion representation rotation is proposed in this thesis,and combining with the research of 3D modeling technology in computer vision,the high precision of 3D model is created and used as reference to determine the pose of a single image.Moreover,The Extended Kalman Filtering(EKF)is used to solve the Perspective-n-Point problem by using quaternion to represent rotation.The main research contents are summarized as follows:The goal of this thesis is to recover the rotation and translation between two views of a moving camera from captured images.The approach proposed in this thesis recovers the rotation and translation changes separately without resorting to an essential matrix.SIFT algorithm is used to extract feature points of the images,and the RANSAC algorithm is used to eliminate incorrect matches.Next,the correctly matched feature pairs are converted from pixel coordinate system to Euclidean coordinate system.In order to construct a polynomial system,the relative rotation of the camera is solved and then the translation vector is recovered by taking the rotation into the equation.The performance accuracy of proposed method is compared with several state-of-the-art algorithms using actual image sequences in the presence of noise in feature point coordinates and for various numbers of feature points.The results show that the proposed approach shows better accuracy in the presence of noise.In order to estimate camera pose of the single image,high-precision 3D model established by 3D reconstruction technology is used as a reference to calibrate the external orientation elements of the single image,in which 3D model is realized through multiple disordered images obtained from different perspectives.It starts by back-projection the 3D models to the image space using a virtual calibration camera with initial parameters,and the image obtained by projection is used as reference image which is matched with query image,registration process consists of two steps: 1)feature extraction;2)similarity measure and matching.Then with the help of the similarity measure,the camera pose of the single query image are estimated by the quaternion method based on the matched feature pairs.Considering that the feature points tracked by the camera tend to drift over time in practical applications.Meanwhile,due to the time dependence and the uncertainty of the characteristics of continuous motion camera pose estimation.In this thesis,Extended Kalman Filtering(EKF)is used to track 3D objects in the sequence image and estimates the position and direction of the camera relative to the scene,which calculates priori estimation of camera pose from the motion model of the camera and then corrects it by minimizing the reprojection error of the reference point.The uncertainty of feature points over time is effectively overcome by this approach.The experimental results show that the presented approach improves the robust of camera pose in the presence of noise.