| With the rapid development of computer technique and network, information format computer can process is not limited in the area of words, table and graphics, etc, and nowadays, the computer can process image, sound and video that have plenty datum and much more complexity.As a new kind of multimedia, video has its own features in expression, processing and management. So research on video becomes one of hottest direction in the area of information. The new international standard of MPEG-4, which is developed and established by MPEG, proposes conception of video object to organize scene. The standard of MPEG-7 aims at describing kinds of multimedia object normatively. But the standard of MPEG-4 and MPEG-7 do not refer to concrete arithmetic in video segmentation, instead, they are left to be researched further as an open part.First, we discuss digital video, analyze its features and research the process of analog-digital transition, and then, we go further research on the model of video segmentation and propose system segmentation model and compare present methods using in video processing.The segmentation of moving objects in video is an difficulty and key point in MPEG-4. The applied value of video segmentation is reflected in two aspects: First, video object can be coded separately to improve coding efficient, so better video picture could be got with lower bandwidth. Second, organizing video content structure with video object can realize video access, alternation, searching and index. Based on these above, this paper mainly discusses the technique that detects moving object from video sequence to make the standard of MPEG-4 and MPEG-7 be applied better.There are 3 steps in video segmentation: first input video, compensate moving and detect scene and segment shot. Then moving object can be detected by the semantic feature and analyze its every thing. At last, we can use this object to do retrieval, code and tracking based on content. If it can not satisfy our demand ,we can process again.However, there is no general and perfect segmentation arithmetic until now, every one is confined to special occasion. This paper aims at still background and segments motion object.In video sequence, the objects that can attract people's attention are generally self-moving, like walking man, moving car and boat, etc. And the background does not change, and even if it changes, the reason is from camera moving or illumination changing. So, an efficient segmentation method can divide the video into self-moving object and still background. The current arithmetic is to segment the moving object in video sequence. So does this paper.First, we should do pretreatment on inputting video sequence to solve the existing problem that lower contrast between foreground and background because of the bad quality picture. So when segmented, parts of object will be lost. Furthermore, many auto segmentation methods are affected by the contrast. To conquer this problem, local contrast enhancement is adopted which mainly use mean and variance in every pixel's neighbor. Because mean and variance are the estimation of average luminance and luminance contrast separately, so threshold method can be adopted to enhance the area where lower contrast occurs. After those processing, little noise would be produced and it happens in segmented pictures in the shape of sole point or point gathering. According to the noise characteristic, 3×3filter is designed to detect and wipe off the noise. After these, video picture contrast is enhanced and the lost of foreground parts is avoided.After pretreatment, the video sequence will be passed through the segmentation system, so motion region can be detected:Considering that the neighbor frame difference consists of two parts: noise and motion area, and in every frame, the distributing of noise and has its own law. Usually, it can be assumed to obey Gauss Distributing. If two random variables both obey Gauss Distributing, their difference also obeys Gauss Distributing, so the relative noise between the neighbor frames obeys Gauss Distributing, too. Thus, through iterative weighting arithmetic, relative noise can be estimated.The rule of choosing weighting function is: According to the difference between every pixel value in difference frame and the mean of relative noise, different weighting value is distributed to every pixel in the next relative noise estimation. The pixel that gets small difference has big possibility to be noise, so the weighting value distributed to it is also bigger. Vice versa.With the estimation, most of noise in the difference frame can be filtered, and then, after being processed further, binary image can be got, but there are a little burrs and holes left. To wipe off these, morphological method should be used.Dilation can combine the background points to the objects. If the distance between the two objects, dilation can connect the two together. After these processing above, the motion region can be detectedBecause precise background is so important for the video segmentation, so next step, we should do Stat on the background--Background Matrix and Counter Matrix are designed, and they have the same size with the video picture. Background Matrix is to store the background information during the process of background Stat, and Background Mask can be refreshed from it. Background counter matrix is to record times that no-changing frames. Through the neighbor frame difference, we do Stat on the background of video sequence, and build credible background region, and make difference with previous frame. And then ,using the frame difference of current frame and previous frame, we analyze 6 things that may occur to estimate motion region and object mask. At last, we do post-treatment to eliminate noise and smooth object borderline and segment the video object.This paper does further research on the video segmentation algorithm, especially on the moving object detection from the static background. The research in this paper has much significance for the techniques of video segmentation in the standard of MPEG-4... |
PDF Full Text Request