Research On Electronic Image Stabilization Based On Feature Optical Flow

Electronic image stabilization technology is a hot research field of computer vision, it can improve the stability and clarity of video camera, and it is widely used in military and civil field. Because the camera system is affected by the attitude of the carrier at different times or the dither, it can lead to the instability and the ambiguity of the video.The essence of the electronic image stabilization technology is the method of image processing. The size of the image offset and the video sequence are obtained. Compared with the mechanical and optical image stabilization technology, it has many advantages, and can significantly improve the stability of the video camera. This paper analyzes the basic principle of the electronic image stabilization algorithm and makes the corresponding improvement from three aspects of motion estimation, motion filtering and motion compensation..Firstly, the camera imaging principle of the system, the establishment of similar transformation model; followed by the analysis of motion estimation, motion filter and motion compensation of the three parts of the classical algorithm, and a simple analysis of the theory, the principle of the proposed algorithm are summarized.The third chapter analyzes the main methods of optical flow estimation, selection of feature optical flow motion estimation and matching; points out the limitations of the traditional Harris feature selection, proposed sub-pixel Harris corner distance limit selection algorithm, by adding the distance limit mechanism, ensure the rationality of the selected corner number, and the uniform distribution of feature points in the image, can improve the speed and accuracy of motion estimation; feature point matching without using the traditional matching method, but use KLT algorithm for feature point matching, can improve the speed of application, light golden pyramid KLT algorithm is only applicable to the shortage of small movement, large motion tracking and matching; finally, using RANSAC method to screen the global motion vector, the global motion parameters is obtained by using least square method.The fourth chapter studies motion filtering and image compensation. The Kalman filtering is applied to the motion separation, and the curve of the random motion parameters is filtered, and the trajectory of the desired motion parameters can be obtained.. In the image compensation, selection of fast nearest neighbor interpolation method, of video image stabilization as compensation; compensation after the image will be at the edge of the black areas, this paper selects FMM fast image restoration technology, stability of video sequences repair, eliminate the black edge, and improve the effect of compensation video observation; finally from the subjective evaluation, peak signal to noise ratio, image difference points of image stabilization performance evaluation.