Real-time Image Stabilization System Using Adaptive Motion Estimation

• Main Authors: Ming-Yu Hsieh, 謝明佑
• Other Authors: Chorng-Yann Su
• Format: Others
• Language: zh-TW
• Published: 2005
• Online Access: http://ndltd.ncl.edu.tw/handle/42138891750185179363

碩士 === 國立臺灣師範大學 === 機電科技研究所 === 93 === In this paper, we propose a new image stabilization system by using adaptive motion estimation. Our aims are to decrease the execution time of a stabilized system and obtain a more stable video sequence. The stabilization system contains two parts: motion estimation and motion correction. In the motion estimation, we use a small diamond search and Sobel edge detection to increase the reliability of the prediction of block motion. And we use an alpha-trimmed mean filter to yield a global motion vector (GMV) for each frame. The proposed filter can remove unusual motion vectors and promote the accuracy of GMV estimation. In motion correction, we make use of wavelet transform. After comparing the usages of different decomposition layers and different wavelet functions, we choose two-layer decomposition and the 9/7 filter function as our experimental standard because of its better performance. For real-time applications, the allocation and the movement of storage memory generally influence the execution time of a system. In order to reduce such time consuming, we store the RGB data in a continuous block of memory instead of storing them separately in three blocks of memory. For the image size of 352 x 288 (CIF format), the proposed system can provide about 20-30 frames per second on average and result in a stabilized video sequence for users. Finally, this system is programmed with Borland C++ Builder.