Research On Unsupervised Video Object Segmentation Method

Video object segmentation is an important branch of pattern recognition and machine vision, which is quite promising in a wide range applications, such as smart surveillance, video retrieval, object recognition, interactive video entertainment and so on. A lot of researches regarding video segmentation have been studied at home and abroad, but many problems still have not been resolved yet. For example, manual annotation object segmentation, ineffectiveness in the automatic segmentation, low efficiency in segmentation and poor robustness issues. The aim of our work is to focus on the automatically-segmenting effectiveness and efficiency in unconstrained video segmentation. The contents and results are as follows:1. The foundations of video object segmentation is introduced in details, including relevant image segmentation algorithms and graph theory. Moreover, the current video segmentation methods and their classification are discussed in a further step. Then, we propose several improved algorithms on the basis of the previous work.2. We propose a novel video object segmentation algorithm based on the combination of LDOF optical flow method and area scan lines. In the algorithm, LDOF optical flow method is utilized to calculate the optical flow of pixels at the difference between tand t+1in a video frame at first. Therefore, rough sketch of video object can be extracted by combining optical flow gradient of a central pixel with directional derivatives of its neighboring pixels. In addition, the outline of moving object is detected precisely and promptly though zone-scanning line method so as to suppress over-segmentation and overcome the shortcoming of inaccurately-extracted object edges from optical flow field. After obtaining the initial outline of video object, the outline model is kept learning while mixed Gaussian function is used to build up color models of foreground and background. In the end, Grab Cut algorithm is utilized to segment video object accurately. The proposed method is able to segment moving objects in videos automatically as well as obtain better effectiveness and efficiency of segmentation.3. We put forward a novel video object segmentation method based on region selection under constrained conditions. In the method, candidate of video object region is established in accord with the total scores from intra-frame and inter-frame. To be specific, using the maximum weight clique algorithm to select the video object region from the candidate region. Thus, the segmentation of video object is recast to deprive an optimal weighted clique from a region map, but mutual exclusion constraint in intra-frame and neighboring constraint in inter-frame are meant to avoid unreliable regions, only to find that the discrete solution can’t be figure out. Therefore, a computing method for weighted clique maximization under a novel constraint is brought up with in the end. Theoretical analysis and experimental results exhibit that our method is accessible in the chapter.