Research On Motion Estimation Algorithms Of A Digital Image Stabilization System

Digital image stabilization (DIS) technique is the new trend of image stabilization which determines the motion vector between frames of a video sequence and compensates it with image processing methods directly, and stabilizes the video sequence, when the carriers of cameras are influenced by irregular jitters which leads to instability of the video sequence. Because of its advantages of high stabilizing precision,compact size,lightweight,low power consumption and reasonable price,the DIS technique has been applied widely in both military and civilian field. In this dissertation, the author mainly focused on the research of DIS technique, especially, paid more attention on the motion estimation algorithms in DIS system. Through analyzing and comparing the characteristics of current commonly used motion estimation algorithms, the gray-scale projection algorithm was selected as the research object of this article, and corresponding improved algorithms were proposed.The main research work in the dissertation is as follows:First, the application background of the DIS was researched, which showed the necessity of image stabilization, and then the basic principles, structure and the key technologies of DIS were introduced. Based on analyzing kinds of motion estimation algorithms, a novel hybrid coarse-to-fine motion estimation algorithm was proposed which combines the gray-scale projection and block matching algorithm. Compared to the traditional gray-scale projection algorithm, the new algorithm introduced the adaptive block matching algorithm for accurate motion vector estimation to avoid the interference of local moving objects. The experimental results showed that the proposed algorithm improved the estimation accuracy of motion estimation by 1 dB compared to the traditional gray-scale projection algorithm, which was close to the accuracy of full search block matching, while the computational complexity increment was not dramatic. The improved algorithm gained good balance between speed and accuracy of the motion estimation.Second, a kind of gray-scale projection algorithm based on log-polar image translation was introduced due to the limitation that the traditional gray-scale projection algorithm could just deal with translational motion, but could not work for estimating the rotational motion vector. When the log-polar transform was used in rotational image, the rotational movements between two successive frames in Cartesian coordinates could be represented by translational movements in log-polar coordinates by using the rotation invariance of the log-polar transform. Accordingly, gray-scale projection algorithm could be used in dealing with the translation motion vector of the two transformed two images. After that, the extracted translation motion vector was transformed to rotation angle in Cartesian coordinates utilizing the inverse transform, with which the current frame image could be compensated to get the stabilized image. Experimental results revealed that the proposed algorithm made the severe rotation in original video improved significantly, and the rotation angle estimation error was less than 1 degree.Third, on the basis of theoretical studies before, a FPGA-based motion estimation system using gray-scale projection algorithm was designed, and static images hardware simulation experiments were implemented. The entire system was divided into three modules according to principles of gray-scale projection algorithm. And pipeline was added to the cross-correlation computation module to increase the processing speed for real-time image stabilization. After initializing the on-chip ROMs with two static images, the timing simulation showed that the computational time of the motion estimation module satisfied the real-time processing requirement, and the gained motion vectors matched with the Matlab simulation results.