Adaptive Algorithm Of Video Object Segmentation Under Moving And Static Background

With the rapid development of the Internet and multimedia technologies, video-based applications have attracted people's attention. Video object segmentation technology as a key aspect of the application of multimedia technology in video encoding,retrieval, multimedia interactive and computer vision has a very important application. Accurate and efficient automatic video segmentation is a key content-based video coding technology. New generation of multimedia standards MPEG-4 and MPEG-7 have adopted content-based video coding standards and video description framework, Including content-based compression, scalability, interactive and several other contents. But the MPEG-4 and MPEG-7 standards only defined the video coding and decoding process and the grammar rule and didn't formulate the concrete video object segmentation method .Video object segmentation as a major supporting technical of video processing technology , video object segmentation study has the profound practical significance and the significant application value. While a lot of research work has been done for the object-based video coding, but so far, a general method has not been found to separate the video object from the scenery effectively. Most of the algorithms are for specific circumstances. Based on this background, This article has chosen the topic which is closely related to the issue.The standards of MPEG-4 and MPEG-7 request that the separated object is meaningful in the semantics. Each part of the same object often has the same attribute. Generally speaking, there are two aspects, namely the time attributes and spatial attributes. they are physical bases of all video object segmentation algorithm.The main performance of time attribute is frame difference, optical flow, or motion vector. The spatial attribute mainly is brightness, color, texture and other transformations or statistical characteristics, for instance gradient image, histogram and so on. Generally, there are two kinds of different consideration angles: One is region and it focuses on the uniformity of spatial attribute. the other is edge and it focus on the difference of spatial attribute. According to the above, we may examine the changes between frames, the movement region as well as the direction and the size of the movement.But by purely using spatial division technology, which involves the using of spatial attributes such as brightness, color or texture, etc,a certain similar regional assembly we get from the video image may often only part of the object. An independent part is often no semantics and is often incompatible with the semantics object. To solve this problem, for the video sequence, we can also use the movement information of the object. In a natural scene, the object movement is steady and not only the object between the neighboring frames has some similarity but also the movement of all parts of the identical object is basically the same. Therefore, if some neighboring regions movements of adjacent frames are the same, we may think they belong to the same object and thus can separate the moving object. We may carry on the video object by using the spatial domain and the time domain information simultaneously. Based on such theory, this paper presents a new video object segmentation algorithm which combined the time domain information and the spatial domain information.Based on the background movement situation, the video sequence may be divided into two kinds of video sequences: the dynamic background and the static background. At present, most video segmentation algorithms are based on a single background. The real-time auto-adaptive static and dynamic background video object segmentation algorithm has become the urgent need. In this article, a new auto-adaptive video object segmentation algorithm for dynamic and static background is presented.First, carry on median filtering and video foreground and background contrast enhancement pre-processing in the pre-stage of the algorithm. Usually, the linear low-pass filter, while removing the noise, will make the clear outline of image fuzzy. But person's vision to the picture edge is very sensitive. Median filter is one non-linear signal processing method and accordingly the median filter is a nonlinear filter. Median filter can overcome the image details fuzzy problem which can be brought by linear filter such as least-mean-square filter, the average filter and so on , retains the image detail well and is most effective to the pulse interference noise and the picture scanning noise. Median filter is particularly suitable for the situation that the image is interfered by very strong noise of mixture of the salt and pepper noise and the impulse noise. Such as the Coast Guard sequence with a strong impulse noise and pepper and salt noise, median filter can achieve very good filtering effect. The edge we get from the image segmentation is often not continual because of the lower contrast of foreground and background in general video sequence. the un-continual edge can not be filled well only through post-processing of the morphology. For the lower contrast of foreground and background, the algorithm adopts the foreground and background contrast enhancement method during preprocessing stage to improve this situation. The algorithm use partial gradient strengthen idea in preprocessing stage. By calculating, only the part of the lower contrast can be enhanced and the higher remain intact. By the contrast enhancement, not only the lower contrast part is strengthened but also the texture of the image is strengthened either and it is very useful for the determination of background displacement.Next, calculate the background displacement after preprocessing. In the background offset calculation, the block-matching algorithm is adopted. First, select a more rich texture block in the current frame background, search for a closer match block with the current frame block in the previous frame, then calculate two frames background offset through the two blocks offset. Certainly, the static background offset is zero and the dynamic background offset is definitely not zero. Then, discover the same part of the two neighboring frames, establish the frame difference model of the same part of the two frames and realize the time domain change detection. Through mathematics morphology processing and scanning processing of the change detection binary template, obtain video object binary template in the time domain. When conducting block-matching calculation, the improved fast three-step search method is adopted .A more perfect matching can achieve under the condition the calculation is nearly same.Then, in the changing regional detecting stage, for some sequences, when changes in the video sequences is static or very small, the video object of some video frame can not be detected or lost. In light of this situation, the real-time calculation and judgment of the mean and variance of the difference template can solve this problem. If the mean and variance of the template is less than a certain threshold, then the video sequence is static and the video object will be lost can be concluded. The previous frame video object output will act as the current video frame video object output to ensure the continuity of the video object segmentation.Finally, the combination of space-time information forms the complete video object. Because the frame-difference model can not determine the accurate boundary of the video object by time-domain information, the combination of the frame-difference model and spatial-domain information can segment the video object more precisely. Carry on the edge examination to the video sequence current frame and complete the video object spatial division. Correcting the change detection using edge information and filling it can get accurate video object binary template.The result shows that the algorithm can quickly and accurately extract video object. The video segmentation algorithm focus on the accuracy, versatility and continuity while minimizing the complexity and computation to adapt to the real-time video communication environment . However, the algorithm is still a lot to be desired. For example, the algorithm can not process the video sequence with the camera rotation. Although the algorithm uses the rapid and simple method as far as possible to reduce the amount of calculation, but also has the certain disparity to real-time picture processing and so on.