Research On Electronic Image Stabilization Based On Motion Estimation And Sepration

Computer vision technology is a hot research field in information science. Througha series of intelligent processing on the video captured by camera, the system can makea reasonable decision to external environment. However, the working effect of camerasystem is influenced by the random motion of the carrier, which leads to the unstabilityand fuzzy of the images. Compared with the mechanical and optical methods, electronicimage stabilization has many superiorities, it can significantly improve the stability andfidelity of the frame sequences, and has a wide range of applications in militaryreconnaissance, traffic, transportation and other fields. This dissertation describes adetailed study on the main algorithms of electronic image stabilization, analyzes thedefect in key links and proposes improved methods accordingly, aiming to improve theaccuracy, instantaneity and integrity.This dissertation firstly outlines the projection principle of camera system, and theexpression of image transformation is also derived; Then gives a brief introduction tothe camera movement characteristics and influence, points out that the global movementof the image background has good consistency; Makes analyses of motion estimation,separation and compensation successively, summarizes the classical algorithms andconcludes their characteristics. Gives a brief introduction of the structure and propertiesof OpenCV, which is chosen as the platform for subsequent research; Indicates the keyissues of the research and proposes some improvement ideas.In the third chapter, the dissertation makes an analysis of the traditional motionestimation methods, the selection and matching tactics of corners; Emphasizes thelimitations of the traditional algorithm, and proposes a new one based on averagebackground method, selection of Harris corner at subpixel level and self-adaptioncontrol of the corner number. The improved method is applied to the experiment, andthe result shows that this algorithm accomplishes higher accuracy and speed, whichensure the motion estimation is reasonable and real-time, and also robust to translation,rotation and zooming; Finally, calculates the motion parameters according to theexpression of image transformation, and extracts the global motion vectors byRANSAC and LSM.The fourth chapter focuses on motion separation and compensation issues. Aimingto deal with the coupling by expected motion and random jitter, Kalman filteringmethod is presented in this paper, to filter the curve of movement parameters. Thesetting of the filter parameters to obtain reliable trajectory is discussed also. Proposes animproved adjacent interpolation method for the motion compensation. Aiming to solvethe problem of the "dark border" after the compensation, a panoramic compensationbased on image mosaicing technique is proposed, which proved can improve the sensory quality and observing effect of output video significantly; Finally, presentsevaluation algorithm of the image stabilization quality, such as PSNR and imagesdifference.