Study On Digital Image Stabilization Technology For Aerial Opto-electric Imaging System

The instability between frames of output image sequence, caused by the factthat aerial opto-electric image system is easily affected by complex environment andplatform change, influences missions like aerial reconnaissance, informationcollection and target tracking seriously. Digital image stabilization, as one ofmethods eliminating video jitter and improving video quality, is essentially a waythat can compensate motion between frames with digital image processing methodso as to reconstruct images and smooth video image sequence. Because of its highaccuracy and flexibility, it is now widely researched and used in both military andcivil affairs. Nowadays the stability of video image sequence is urgently required byaerial opto-electric system, so this dissertation mainly focuses on digital imagestabilization and developed several efficient image stabilization algorithms.Digital image stabilization technology as primary topic, its basic principle,structure and key technologies are researched systematically. Some typicalalgorithms that are commonly used like global motion estimation, motion filtering,motion compensation and image compensation are analyzed and compared.According to actual application, three digital image stabilization algorithms aredesigned so as to compensate video jitter of aerial opto-electric imaging system indifferent motion patterns. This dissertation’s main research and results are as follows:1. Digital image stabilization is analyzed in detail. The impact of camera jitteron image is analyzed first. Then basic imaging principle and motion patterns ofcameras are introduced and homography matrix which is commonly used in globalmotion estimation is derived. At last, the coordinate conversion model of2D imagemotion is described.2. For video sequences with panning jitters, the impact of large local motion orsignificant illumination change on accuracy of global motion vector estimation ishard to be overcome by traditional gray projection algorithm. For projection curvewith large difference, gray parts are used as matching units and Dynamic TimeWarping (DTW) algorithm is used to deal with incorrect projection curves so as toget optimal deformation path. Experiment shows that the proposed algorithm canreduce occurrence of wrong match and improve stabilization effect for videos withlarge local motion and significant illumination.3. For video with intention movement, an algorithm based on Hilbert-HuangTransform (HHT) is proposed. First global motion vector is extracted by Harriscorner detection, composed of unwanted jitter motion and intention motion. Thenempirical mode decomposition (EMD) is imposed on global motion vectors to getlimited intrinsic mode functions (IMF) with frequencies from high to low. At last,Hilbert transform is imposed on each IMF to get Hilbert time-frequency-energyspectrum and distinguish low-level IMFs with lower energy and high-level IMFswith higher energy so as to separate jitter motion and intention motion effectively.Experiment shows that this proposed method can eliminate high frequency jitterefficiently and get smooth videos.4. For situation when long-time fast stabilization is needed, an algorithm basedon fast smoothing point-feature trajectories is proposed. First, the inherent defects oftraditional image stabilization models are pointed out, based on which a systematicframework mainly containing modules getting and smoothing feature trajectories isbuilt. Then local image features are specially researched. An improved SURF algorithm is introduced to extract feature points from original shaky video andDelaunay triangulation algorithm is used to determine adjacency of these featurepoints and generate point feature trajectories. At last, an objective function takingboth of feature smoothness and video quality degeneration into account to smooththe feature trajectories and get stabilized video. Experiment shows that the proposedmethod can prevent information loss and eliminate accumulated matching errorbetween frames in motion estimation. So it can be used in long-time imagestabilization.