| 英文摘要 | People can easily acquire huge videos as a result of the development of computer technology, Internet Technology and Manufacturing technology of all kinds of cameras. For it contain abundant information in videos, video analysis technology is given more and more attention, such as video tracking, fluid analysis, 3D reconstruction, etc. In many other fields of science where need image sequences to conduct the researches, such as oceanography, atmospheric science, climatology, medicine, people also have done a lot of theory and application works. As an important aspect of video analysis, optical flow estimation is becoming popular among researchers, because of its advantages such as the direct description about motions and not easily affected by the appearances of motion objects. Now, optical flow estimation is one important method in fields of motion detection, edge extraction and others. Though great achievements have been done about motion estimation and it can be applied to many aspects, there is still much work to do. In this master thesis, the theme is to retrieve more accurate motion field in videos. The specific objectives of this thesis are to 1. build more descriptive optical flow estimation model, 2. build model that can retrieve more accurate motion field for multi-motion and segment the motion simultaneously. Based on a general review and analysis of existing methods on motion estimation (Chapter 2) and after studying the above two objectives, we propose two optical flow estimate methods. The main contributions are, 1. Propose Statistical Modeling of Optical Flow method. In this method, by introducing Radial Basis Function Neural Network model, the statistical for brightness variant and spatial derivative of velocity field are learnt, and thus establish the statistical model of optical flow. The advantage of this method is it can be applied to more motions by choosing different training set. 2. Propose Optical Flow Estimation and Segmentation in Image Sequences. In order to conquer the problem in existing methods which can only retrieve motion field for one specific type of motion, and thus not robust for multi-motion, we propose motion regularization selection method. It is a more general method. By propose basis motions and combining them efficiently, the method can choose the best basis motion combination for different motions. And thus can we retrieve more accurate motion field. The contributions of this method are: 1. For different m... |
修改评论