Research On Video Object Segmentation Algorithm

Content-based video object segmentation has become a hot topic in the digital technology, also the computer vision. The segmentation of moving objects in video sequences is very important for second generation coding and is a basic prerequisite for content-based video applications, which contain video searches, compression and object-oriented edition, aptitude human-machine exchange, etc. The results of objects segmentation will affect subsequent applications directly. At the present time, there is no current method, which can segment object models from the background efficiently, though a great deal of research work has been done for video coding.Based on research on methods for the segmentation of moving objects in video sequences, two new methods are proposed to achieve automatic segmentation from video sequences for the problem of low anti-noise, poor accuracy and shadow elimination.(1) Video Object Segmentation Based on SVMFirst, the video is divided into background part and target part in essence, and the segmentation problem is regarded as classification. Each pixel is determined for target or background, so it has no influence by tensile, translation and the rotary of the lens. The mask of the training sample of the SVM is extracted by fourth-order statistics detection method based on the gray differences between multiply frames. Then segment the frame using the airspace segmentation algorithm, and time-space combination to get the training sample, and then extracting the feature, training, get the SVM, using the SVM to segmentation of the fore and background, getting the mask of the object.Then the space segmentation is got by watershed algorithm.At last, combined the SVM mask and the space result, and the object is got.Experiment show this method has anti-noise to some extent, and improves the accuracy.(2) Video object segmentation based on background construction First, moving regions are achieved through temporal segmentation, getting rid of making frame difference or twice frame difference image as models. Frame mask image can be achieved with a threshold from moving information got by intersections of twice frame difference images.Then background frame is initialized by a certain algorithm, in which, a counter is used for accumulation, and makes the constructed background more reliable. At last, the moving object is segmented by current frame difference from predicted background. And a new method is proposed to solve the shadow elimination. Experiment show it can solve the shadow elimination effectively.