基於影光流動技術的數位影像穩定系統

Abstract

本論文提出一種以影光流動方向(optical flow)為基礎的數位動態影像穩定系統。系統可分為三個步驟 1) 運動估測、2) 估測向量的抽取、3) 影像運動修正。首先，在運動估測中我們使用疊代的影光流動方向 (optical flow) 演算法在每二相鄰影像序列中得到一組局部影光流動向量，且每一像素皆有一對應的流動向量，但其中可能包含一些不準確的影像運動估測，因為輸入影像本身可能有誤差; 接著在向量抽取步驟則應用Hough transform在水平移動量和轉動量的參數空間中選取最大可能的運動估測量。最後，在影像運動修正步驟中使用運動估測量去計算此運動近似的振動頻率，並據以估測原來影像的運動而得知所需的反饋向量，進行反向運動修正，使穩定後的影像序列能達到清晰穩定的效果。

This thesis proposes a digital image sequence stabilization system based on optical flow techniques. The system includes three major modules 1) motion estimation, 2) motion extraction, and 3) motion correction. First, in motion estimation module, we employ the optical flow technique to obtain the local motion vector field. The outputs of optical flow estimation are the velocity vectors corresponding to every pixel in the current image frame. There could be many inaccurate motion estimations in the motion estimation procedure because optical flow techniques are rather noise sensitive. The motion extraction module makes use of a new version of Hough transform, which acquires field motion vector in terms of parameter space of the image field. The motion parameters to be estimated for the current image frame are angular velocity, and horizontal and vertical translational motion relative to the previous frame. Then the motion extraction module plays the role of extracting the most possible motion parameters. The motion estimation module can produce very precise field motion vectors. Finally, the correction module integrates the field motion vectors and performs the de-motion of correcting the residue shake motion, including translation and rotation as well.