The Motion Segmentation Based On DIRECT

Motion segmentation is an important research in computer vision. it is a basic application research and is challenging. It is not limited to some special problems, but used in various subject. The research has important theoretical value and practical significance. Many years, with the continuous study and practice, as well as the rapid development of multimedia technology, it has been made a great progress on the research of motion segmentation, and has been applied on many areas. But so far on the subject there are still many issues unresolved.In this paper at first, we make a simple comment on the characteristics of the common methods in motion segmentation, and then focus on three aspects of the approaches used in this paper:First, determine an initial changing area using a method based on HOS algorithm. In this section, we at first introduced motion detection and the fame model, then the HOS algorithm which is a kind of statistical methods is used to detect the changing area. As for the solution of the noise variance in this paper, we use an iterative method to estimate the noise parameter. Finally this parameter is used as the threshold of the HOS algorithms.The second aspect is the motion estimate based on parameter model. In this section, we use the axis affine model as our motion model in this paper, and a global optimal algorithm with bound constraints--DIRECT algorithm as our methods of motion parameters estimating. At the estimating process in this paper, according to the principle of absoult difference, we define our objective function, and determine the supporting region of each moving object on the detecing area. Then we use the DIRECT algorithm to estimate the axis affine parameters, and convert to the traditional affine parameters. This approach is built on the equivalent relations between the axis affine model and the traditional affine model, so it can only use a small quantity of pixels to get the result quickly. Meanwhile it makes the geometric meaning of motion parameters clear, so we can easily determine the changing bound of the parameters.Third, the motion segmentation based on MRF model. In this section, we at first introduce the basic theory of MRF, and then compute absolute difference between the pixels on which one pixel belongs to the current frame, the other one calculated through the affine parameters belongs to the next frame. This absolute difference is used as our data term of energy function. Finally we use the expansion moving method which is based on the graph theory to minimize MRF energy function. In the experiment, it proved that this method can get a much more accurate segmentation result compared to the ICM method.