Video Motion Segmentation Based On Dense Optical Flow

Video motion segmentation has been the basic research branch in the field of computer vision. It is the basis of the intelligent monitoring, human-computer interaction, navigation and guidance, industrial robots and other applications. Video motion segmentation refers to dividing the target area from the video or image sequence according to certain criteria such like the differences on the edge, texture, spatial and temporal characteristics and other aspects of the moving object. However, in practice, some factors, like complex scenes, the camera moving, multi-target or blocking, make it very difficult to make a precise and stable motion segmentation.Based on deep research of video moving object segmentation algorithm that has been proposed, this paper presents a simple and effective segmentation algorithm. The main work is as follows:(1) The basic concepts of image segmentation and motion video object segmentation are introduced, and the division standards in different video motion segmentation algorithms are analyzed as well as their advantages and disadvantages. We construct a feature description based on spatiotemporal because it is difficult to describe the information of the moving object in the video using some simple characteristics like edge, color and texture.(2) In this paper, a dense sampling of the video frame is carried out in order to get enough exercise information. And these sampling points with temporal characteristics are filtered to remove some points with unobvious structural information that are difficult to track. Then the optical flow method is used to track and extract trajectory form the remaining sampling points. To describe the motion information of the moving object, a concept is defined, which is the distance between two trajectories within a specified window of time.(3) A video motion segmentation algorithm based on spectral clustering is proposed. Using the phenomenon that the trajectories of the moving objects in the video are similar, the trajectories of the dense sampling points are clustered to achieve segmentation of moving objects. The similarity matrix is constructed based on the feature that is the distance between trajectories. Besides, what proposed is adding the structure information of the moving objects in the video to improve the effect of clustering while using the classical clustering algorithm for clustering the similarity matrixes, the feasibility of which is verified by experiment.