Research On Image Correspondence Techniques And Applications Based On Phase Information

Image correspondence is one of the basic technologies in the field of image processing and machine vision. The purpose of the image correspondence is the estimation of the transformation between the image pairs, which aligns the image pairs optimally in the the space. With the requirments of the intelligence system for military and civil use, image correspondence technique has developed rapidly. However, it is still an open problem to get robost, high accuracy and high efficiency image correspondence, espically the image correspondence in a complicated scene.This dissertation researches the technologies to deal with difficulties in the image correspondence, such as robust image correspondence technology in the structural environment, high accuracy image correspondence technology, robust image correspondence method in the complicated unstructral environment using phase information of image, and the application of image correspondence method based on phase information of image on robot's monocular odometery. The main contents of this dissertation include the following aspects:A novel shape representation and registration method of rigid transform based on correspondence Hough spectrum is proposed. The Correspondence Hough spectrum (CHS) is defined for the correspondence and representation of arbitrary shapes, and the translation, rotation, scale and periodic properties of CHS are summarized and proved as well. The 2D image is transformed to a 1D Hough spectrum. The rotation and scale parameters are decoupled with translation, the rotation is transformed to the phase translation of the Hough spectrum, and the scale is transformed to the scale of amplitude. Then the rotation and translation parameters can be estimated easily from the phase and amplitude of 1D CHS, and the parameter of translation can be estimated in the Hough density space. Experiments show that the method is effective to estimate the parameters of translation, rotation and scale, and is immune to noises and overlapping also.An accurate image registration method based on Enhanced Phase Only Correlation (EPOC) is developed. The EPOC uses a hierarchical strategy to estimate more accurate image pair's registration parameters, which consists of a coarse estimation and a robust and efficient refinement stage as well. The initial parameter is estimated through a conventional Phase Only Correlation (POC) method in the coarse stage, and then it is refined by the local upsampling Fourier transform in frequency domain to achieve higher accuracy. Furthermore, as will be shown in many experiments, the EPOC can achieve more accurate translation and rotation estimation, and it is efficient and robust to noise.To the image correspondence of complicated scene, a multi-scale optimal function method based on phase coherence in hyper-complex wavelet domain is proposed to estimate nonlinear image motion. Based on the 2D Hilbert transform and quaternion definition of analytic signal, a hyper-complex wavelet is constructed, whose phases encode the vertical and horizontal translation of image locally. And we construct an object function of local phase difference, whose extremum is the constraint of local image motion. Then the multi-scale phase unwrapping strategy is used to estimate the image motion of small scale from the larger scale. Lots of experiments have shown that hyper-complex wavelet phase coherence and unwrapping based optical flow estimateion method developed in this paper is more precise than Wai Lam Chan's method, and it can estimate the image motion of complicated scene efficiently.The application of image correspondence algorithm for a visual odometry strategy using only one camera in country road is proposed. The multi-scale optical flow computed by an efficient hyper-complex wavelet phase coherence method has been used to represent the distributed image motion derived from vehicle motion, which can robustly detect the image micro motion through multi-resolution image signal phase at scatted regions even if there are few regular textures in country road. After analysis of robot's motion characteristic in country road, a 4-DOF ego-motion model is proposed to describe the relationship between image motion and robot motion. The outliers in the optical flow field are removed by a blocked RANSAC algorithm, and the pitch compensation algorithm is used to weaken the ego-motion error induced by pitch angle. Experiments have shown that the hyper-complex wavelet phase coherence and unwrapping based optical flow estimation, 4-DOF ego-motion model and the pitch compensation algorithm are effective to estimate robot's ego-motion parameters in country road.