Image Registration Based On Feature Point And Its Application To Electronic Image Stabilization

Image registration is a very important technique in aerial remote sensing and stillhas many unsolved problems. Its main purpose is to realize the process of geometricallyoverlaying two or more images of the same taken at different times, from differentviewpoints or by different sensors. Image registration is a crucial step in many imageanalysis tasks which include change detecton, information fusion, environmentsurveillance, image mosaic, object recognition and so on. Image registration still hasmany unsolved problems. Firstly, the ambiguity caused by dynamical objects,illumination change, large geometric transformation, similar patterns and lowoverlapping area between images can decrease the correct matching points and increasethe mismatching points. Secondly, registration of images with complex non-linear,locally dependent geometric distortions, multimodality and3-D images belong to themost challenging tasks at this moment. To solve these problems, we are exploring threebasic problems-how to exact accurate feature pints, match control points and whatmapping functions to use for accurate registration. At the same time, the research resultsof image registration is applied to video stabilization in dynamic sequence, which canimprove the stability and fidelity of video sequence recorded from cameras on movingcarriers.In this dissertation, the effect of corner based feauture point detection, descriptorbased feature point matching and feature spatial relation on image registration is firstlyanalyzed across-the-board and depicted quantificationally, and then some feature-basedimage registration methods and registration-based stabilization are developed．Finally，the proposed methods are vertified by a lot of real aerial image experiments．The mainresearch work and contributions of the dissertation are listed as following：1. Traditional feature point detector algorithms computer the image gradient based ondiscrete pixel differences, and finite differences can provide a very poorapproximation to a derivative, a new optimal derivative filters based point detectionalgorithm is proposed. The basic principles of several first-order and second-ordergradient based feature point detectors is sdudied systematically, including Harrisdetector, Hessian detector, Harris_laplace, Hessian_laplace. The performances ofthe above point detectors when applied in feature point detecting are evaluated. Thescope of application and deficiency for each point detector are also analyzed.Optimal derivative filters method in general has emerged as an optimization of therotation-invariance of the gradient, we extend the application of optimal derivative filters to realize the accurate locations of the corners. To ensure that the samefeature points are detected in images with different focus from camera, we choosefeature points that constant across the three resolutions. Then minimum spanningtrees (MSTs) is used to find initial matching. Once matching feature points hasbeen found, the transformation parameters which has global minimum error arethen estimated using non-linear squares (NLLS) and random sample consensus(RANSAC) method, and Finally the image registration was finished. Experimentresults show that by accurate registering frames at the background using optimalderivative filter and globally-optimal transformation model, the technique arerobust under different dynamic scene and illumination.2. The classical way of comparing two moment descriptors only takes into account themagnitude of the moments and loses the phase information. The novelty of ourapproach is to take advantage of the phase information and Magnitude informationin the image registration process while still preserving the invariance to rotation.Firstly, interest points in an image are detected by Harris-laplace operator, themoments defined on the scale normalized interest point neighborhood werecomputed; And then using the magnitude and phase angle information of moments,through comparing the Euclidean distance of these moments to extract the initialfeature points pair; after the geometric transform of input images, the imageregistration was finished. The experimental results show that the proposedalgorithm is robust to scaling, rotation, noise and effectively reduce the false matchrate, realize image registration accurately.3. The Harris detector can produce false and unstable corners, and the matching pointshave different accuracy, an aerial video registration algorithm using projectiveinvariant is proposed. This method can separate object motion from camera motionin an aerial video, consecutive frames are registered at their planar background.Firstly, the Delaunay Triangulation is used to find initial matching. Then the most“useful” matching points that best satisfy the cross-ratio invariant are presented toestimate the geometry transformation, and the image registration was finished.Experiment results show that by accurate registering frames at the backgroundusing Delaunay Triangulation and cross-ratio invariant, a method is provided todetect the moving objects.4. Due to increased image size and geometric difference between multi-view images,the image registrations become complex. We present a new method to handle thisdifficult problem. Various steps are built into the method to detect and remove the incorrect and inaccurate correspondences. The matching points that satisfy theprojective invariant are used to update the triangulation and register the images in apiecewise manner. Image subdivision reduces the geometric difference betweenregions that are registered and simplifies the correspondence process.5. In-depth researches the the key technology of electronic image stabilization and weproposed aerial video stabilization algorithm based on3D motion model. Firstly,we introduce the development of image stabilization and electronic imagestabilization technology, basic principle and mathematical model of electronicimage stabilization. Then motion parameters are determine between two imageusing3D motion model. Finally, we have focused on improved Kalman filter basedmotion compensation methods, and we propose an adaption Kalman filter, whichcan solve the problems of how to extract scanning motion vector from globalmotion vector and compensate the current frame quickly. The experimental resultsillustrate that the proposed algorithm can stabilize the inter-frame jitter and trackthe real scene effectively.