Research On Content-based Video Scene Analysis

With the development of multimedia and network technologies, a growing number of video applications have emerged. Therefore, how to manage and retrieval these data efficiently is going to be an emergent topic. The key point of resolving this topic is hierarchical analysis based on video structure assumption. During the past decade, video shot segmentation technology has been well researched; video scene analysis is going to be a more challenged point. Video scene is the minimal semantics cell in video structure, so the development of video scene analysis technology could be able to boost the process of multimedia research and application.The main technologies of video scene analysis, especially low-level feature, video shot segmentation and distance measures are introduced. This paper analyzes the methods on video scene analysis algorithms, and abstracts the general framework of resolving the problem of video scene analysis.Then, a new method for Scene Detection is proposed, which based on local feature and scene feature complement with each advantage and frequent shot set. This method calculates shot similarity matrix using GIST feature from shot key frames which are typically selected, and frequent shots are selected from the key frame sequence using column-ranked shot similarity matrix. A visual vocabulary is created from local descriptors of these key frames resulting to a visual words histogram representation for each shot. The localization of scene boundary stage used histogram of visual words by calculating the similarity between each homeless shot and its pre-and post- frequent shot set, and assign it to the closer. Extensive experiments are tested on three movies.Lastly, some experiments have been conducted on analyzing the effect of different features and stages of the algorithm based on frequent shot set in the task of video scene segmentation.