Fast Video Stabilization Using Principal Axis Analysis

• Main Authors: Chin-Wei Hsu, 許沁瑋
• Other Authors: Shyi-Chyi Cheng
• Format: Others
• Language: zh-TW
• Published: 2010
• Online Access: http://ndltd.ncl.edu.tw/handle/50990582064742239877

碩士 === 國立臺灣海洋大學 === 資訊工程學系 === 98 === Digital image stabilization uses image processing and computer vision technology to reduce unnecessary vibration and smooth instable video sequences. There are two sources of motion in an instable video sequences: the first is from the moving objects, such as the moving car; the second is caused by camera shake, such as hand shakes of photographer. The problem of video stabilization is to eliminate the second motion, which leads to high computational complexity. How to speed up the processing with good image quality is an important issue. In general, digital image stabilization consists of two parts: motion estimation and motion compensation. Motion estimation calculates the global motion vector between two consecutive frames. Excluding the motion vectors of foreground objects, the global motion vector caused from camera jitter is regressed from local motion vectors. Then motion compensation uses the compensation motion vectors which are obtained from the global motion vectors of video frames to reduce the uncertain camera jitter and stabilize video sequences. In the past, most of the image stabilization techniques are applied to video sequences with fixed or panning shooting. In this thesis, we present a fast global motion estimation algorithm for video stabilization using principal axis analysis, which solves the slow computation problem in existing motion estimation. The proposed method calculates invariant principal axes in backgrounds from consecutive frames to obtain the global motion vector considering translation, rotation, and scaling factors. The global motion vector is then used to calculate the compensation vectors to eliminate un-expected jitter on video sequences. Experimental results show that the proposed method is faster than the compared method.