Research On Unsupervised Object Segmentation In Videos

The emergence of video massive data in recent years, this brings greater challenges to the video storage and video content analysis. In the video content analysis, the object is usually attracted our attention and the focus of our research, but how to segment the interested object from the background is an important topic in computer vision, and has gained great research and application values, especially plays an important role in the field of pose recognition, video retrieval and video summaries. However, how to automatically discover the object of interested in the video, and accurate segmentation of the object is a very challenging problem. Meanwhile, in order to meet the demand of large video data rapid processing, it is necessary to achieve a fast segmentation method. Therefore, we proposed two different unsupervised online object segmentation algorithm which fused object appearance features and motion features, as well as given an accelerated algorithm to achieve a rapid object segmentation.The main contribution and innovation of this thesis are as follows:1) A automatic object segmentation method was proposed, which fused appearance and motion features. The object is initialized by using of the motion boundary which generated by the motion, and the appearance boundary of the object edge feature. Then estimated gaussian mixture model respectively of object and background. A Markov random field model was constructed by taking the superpixels as nodes, and integrating the appearance model and the location prior, also combined with the spatial-temporal smoothness, and gained segmentation result by graph cut.2) A salient object segmentation method based on edge-preserving filtering was proposed. The salient object discovery is formulated as an energy minimization problem, which fuses the appearance and motion features. The edge-preserving filter maintain the objective well boundary information, thereby obtained more accurate object initialization results, then gained object segmentation result accurately.3) A fast online object segmentation method, which based on sub-sampling with gradient-driven and up-sampling with CLMF edge-preserving filtering of the video frame, and achieve an online object segmentation. This method preserves the maximum gradient information in sliding window to down-sampling of the video frames, then use of object segmentation algorithm to segment the object, and local multi-point crossover filter under the guidance of the original video frame, preferably protecting of the boundary information of the object, and greatly improving the segmentation efficiency. As well as, use the object initialization of the current frame to optimize the appearance object Gaussian mixture model, achieved an off-line video object segmentation.