Research Of Video Stabilization Technique Based On Dynamic Displacement Field Model

As is known, video shot with a digital camera usually suffers from variousdistortions due to unstable random camera motion, which affects the quality of theobtained image sequence seriously. For video manufacturing, these image distortionsaffects the quality of the video program greatly; For video tracking systems, imagesequence distortion may create a tracking error or, in some cases, even leads to failureof tracking; For military application, these undesired fluctuations will affect thetracking accuracy or guidance accuracy. Consequently, the stabilization of the imagepickup systems under the unstable or moving circumstances becomes very important.Video stabilization technique removes those undesired fluctuations, estimates theintentional global motions of the camera and compensates them, so that we canremove or reduce those jitter fluctuations' influence and obtain a clear and stabilizedimage sequence. Video stabilization technique has been used widely not only in thecommercial application, but also in the industrial and military areas. The techniqueincludes modeling and estimating the camera motion, filtering the undesiredfluctuations, evaluating the stabilization's performance, and so forth.This dissertation is dedicated to discussing the following several primarytechniques involved in the video stabilization technique: the modeling technique ofdistorted image sequences, the stabilization technique based on the spatiallyinvariable distortion model, the stabilization technique based on the spatially variabledistortion model, the motion state estimation of moving targets, the identificationtechnique of the image distortion model and distortion frequency, the evaluationtechnique of the stabilization's performance. The innovative points in this dissertationcan be listed as follows:1. Based on the 2-D dynamic displacement field model, we extend the dynamicdisplacement model from 2-D to 3-D. It realizes the description for the motion state estimation of the target in the depth direction of the sight field.2. Based on the 2-D spatially invariable dynamic displacement field model, a motionfiltering technique for spatially invariable image sequence is proposed, which isaccording to the relationship connecting the jitter frequency and the motion model.This filtering technique realizes the motion estimation, motion smoothing, andmotion compensation.3. Based on the 2-D spatially variable dynamic displacement field model and themotion filtering technique of the spatially invariable image sequence, a techniqueis proposed for stabilizing the projective distorted image sequences by means ofspatially variable sequence's motion estimation and motion filtering techniques.4. Based on the 3-D dynamic displacement field model, it has been implemented tostabilize the tracking sequence and estimate the target's motion state.5. Based on the estimated non-stationary displacement series, the distortion modeland distortion frequency of the real-world image sequence are identified bymaking use of the Hilbert-Hung Transform (HHT) technique.6. Based on the visual interests quality assessment method of image, a newevaluation of the stabilization's performance is proposed, which combines theobjective evaluation method and subjective evaluation method. Compared it withthe objective-fidelity-only evaluation, the new evaluation reflects thestabilization's performance more effectively.