Study On Algorithm Of High-accuracy Image Sequence Stabilization Using Feature Registration

Currently, the technology of EIS is an important research area of video processing and computer vision. Because of its advantages of high image stabilizing precision, compact size, lightweight, low power consumption and reasonable price, it will have extensive application prospect in military affairs and civil cameras. The key technology of the EIS includes motion estimation and motion compensation, and the motion estimation between adjacent frames is the foundation of EIS. Thus, a lot of research are carried out and several algorithms of motion estimation are proposed. Basing on the analysis on advantages and shortcomings of the typical algorithms, the key technology of EIS is discussed and an algorithm based on feature registration is presented. The proposed algorithm is both of academic value and practical value. The main research works in this thesis are as follows:①The research progress at home and abroad and the common methods of EIS are reviewed and summarized. In this thesis, the basic principles, methods and process of the electronic image stabilization are described, and the key technology of the realization is discussed. Several typical algorithms of EIS and their advantages and shortcomings are introduced and analyzed. It emphatically points out that the algorithm based on feature matching is considered as the important method, because of its perfect performance to deal with the complex dither between image frames. And this kind of algorithm is used in the thesis.②The object of motion estimation is to analyze the adjacent frames and estimate the motion parameter of the camera. How to improve the efficiency and accuracy of the motion estimation is the research emphasis. In order to keep the video stream stable when the camera works in complex movement state, a high-accuracy electronic image stabilization method based on edge image registration is presented. In the proposed method, edge feature and corner are detected from the frames at the same time. Matched points are then utilized to guide edge matching, and a proper number of well distributed corresponding points are selected to get initial global motion estimation. Finally, to obtain accurate global motion parameters, improved particle filter is applied to edge image registration to further improve the image stabilization precision.③When the movement of the camera include random jitter and intentional motion, we must distinguish them firstly. According to the basic principle of Kalman filter and the mathematical model proposed in the thesis, a Kalman filter used to distinguish camera scan from dithering by smoothing the global motion vector are designed. Experimental results show that the method can not only alleviate the vibration but also can preserve the initiative scan of the camera.④Image tailoring and mosaic are often used to reconstruct the undefined regions which generated after the dither image compensated. The thesis analyses the respective characteristics of them. And the image mosaic is used to reconstruct the undefined regions. This method solves the common problems of information loss and image degradation brought by the undefined regions, and also improves the visual effect of the image sequence.Evaluation indexes of the EIS are introduced at the end of the thesis and several image stabilization experiments have been carried out with the video image sequence taken by hand-held camera. In addition, the stabilization effect is evaluated. The simulation and experimental rusults demonstrated that this method of EIS proposed in the thesis is feasible.