Research On Video Segmentation And Merging Based On Global Trajectory Analyze

With the popularization of portable cameras like smart phones and digital cameras, there is an explosive growth in Internet video data. These videos enriched the spiritual life of people, but posed imperative question that how can we do effective edition, or extract out useful information on these videos. Particularly, quick and accurately segmentation of object regions from a video sequence, and merging them naturally, authentically into another video, making the video more interesting, come into a heavy demand in fields like multimedia editing, digital entertainment, film and television production, video coding, security monitoring, intelligent video analysis and video retrieval. Recently video segmentation and merging are mainly be done frame by frame artificially, which takes a lot of time and manpower. As a consequence, video segmentation and merging have come into a hotspot of computer vision.In computer vision both of video segmentation and merging are basic researches with a lot of outstanding achievements. However, due to the complexity of environment and shooting conditions of videos, especially for mobile camera, this problem remains unresolved. In this dissertation, we introduce the basic conceptions and theories of video segmentation and merging, and their research status and mainstream algorithms. Base on that, we improve the traditional trajectory analyzing algorithm, and introduce it into video segmentation framework, proposing a novel moving object detection and an interactive video segmentation algorithm. What’s more, based on trajectory analyzing, this dissertation also propose a novel video merging algorithm, which can quickly, accurately and truly merge an object into a video that is taken in another situation. In this dissertation, the research content and innovation points are as follow:1. We propose a global trajectory analyzing based moving objects automatic segmentation algorithm. Traditional long trajectory analysis cause large occluded area and abandoned its motion information, eventually led to difficulty in boundary determination between targets and background. By designing trajectory distance measurement, this dissertation puts forward a short trajectory utilization scheme. At the same time, Segmentation algorithm based on superpixel is proposed to improve the accuracy of the foreground and background labeling step. Experiments indicates that the algorithm shows great accuracy and stability compared with traditional method.2. We study the traditional interactive video segmentation algorithm, and propose a global trajectory analysis based interactive video segmentation algorithm. Traditional interactive video segmentation algorithm only use the relationship between corresponding pixels between two frames to spread the confidence probability over the video, which ignore the global motion, and leads to serious error in complex environment. If a frame is labeled into foreground and background, feature points in it are also classified. Based on this classification, we propose a novel interactive video segmentation algorithm based on global trajectory analysis. Experiments indicates that the algorithm is robust even in very complex environment and strenuous motion situations.3. We propose a video merging algorithm based on Homography transformation. Traditional video merging algorithms tend to ignore the relative motion between target and background, which leads to "drift" between them, which debased the visual Performance of merged video. This dissertation propose a Homography based motion estimation and compensation algorithm. Experiments shows that our algorithm can eliminate background "drift" phenomenon in merged videos.