Robust Global Motion Estimation and Its Video Applications

• Main Authors: Chang Hung-Chang, 張宏彰
• Other Authors: Shang-Hong Lai
• Format: Others
• Language: zh-TW
• Published: 2004
• Online Access: http://ndltd.ncl.edu.tw/handle/15339143308491195101

碩士 === 國立清華大學 === 資訊工程學系 === 92 === Many applications require the extraction of some basic information from a video for processing, analysis or retrieval. Global motion estimation is popularly demanded in various video applications, including video mosaicing, video stabilization, moving object segmentation and video retrieval. In this thesis, we propose two new and robust global motion estimation algorithms; namely the trimmed least-squares global motion estimation algorithm and the long-term shortest-path dominant motion estimation algorithm. For the first algorithm, we apply trimmed least-squares estimation to fit the computed optical flow vectors to an affine motion model and reject outliers by discarding the optical flow vectors that contain large model fitting errors. For the second algorithm, we first apply the RANSAC method to find multiple dominant motions for every two adjacent frames in the video and then find the long-term dominant motion trajectories by solving the shortest-path problems with Dijkstra’s algorithm. Video stabilization and camera motion classification are the two focused applications based on the proposed two global motion estimation algorithms. For video stabilization, we employ the proposed trimmed least-squares affine motion estimation algorithm to compute the simplified affine motion parameters and then apply the regularization-based smoothing method to these parameters. Experimental results show that we can obtain a stabilized video in real-time speed. For the camera motion type classification, we first estimate the most dominant motion trajectory by the proposed long-term shortest-path dominant motion estimation algorithm and then classify the camera motion type by using a feedforward neural network. Experimental results show accurate classification results on various types of camera motions.