【作者】 周斌；

【导师】 陈贺新；

【作者基本信息】 吉林大学 ， 通信与信息系统， 2008， 硕士

【摘要】 随着CBVAR技术的发展,体育视频中事件的自动提取和检测的研究引起了很多关注。足球是世界上最广泛的体育运动之一,深受广大球迷的喜爱。通过研究发现,足球视频每场比赛持续的时间比较长,但其中真正能够吸引观众注意力的只是很少的一些精彩镜头。如果能够从足球视频中自动地探测出这些精彩镜头,将对电视新闻的制作、视频数据的检索、交互式电视转播以及辅助训练等方面产生深远的影响。所以,本文选择足球视频作为主要的研究方向。镜头检测是基于内容的视频分析和检索中的一项最基本和最重要的内容,本文在介绍数字视频的基本知识和视频压缩的基本理论的基础上,讨论了镜头变换的类型及其表现方式。对压缩域和非压缩域的几种常用的镜头检测方法做了详细分析,通过实验对这些方法进行了比较。本文采用了基于宏块类型信息的方法,这种方法不仅可以有效的检测出突变镜头和部分渐变镜头,而且处理速度远远快于非压缩域中的方法,比较适合足球视频的镜头检测。在基于内容的视频检索中,视频被分割成镜头之后,为了建立视频索引或视频摘要,往往需要提取镜头的关键帧,用来描述一个镜头的基本内容,关键帧提取在基于内容的视频分析和检索中具有重要的作用。本文提出了一种基于帧统计的关键帧提取方法,可以简单有效的提取关键帧,基本上表达了镜头的含义。接下来介绍了足球视频分析的关键技术——特征提取,讨论了几种典型的特征提取算法。并结合足球视频自身的特点,提出了一种基于帧图像局部特征的镜头分类算法。此算法设定了两个阈值,通过阈值的改变,用户可以快速有效的找到自己感兴趣的视频内容。并且根据实验结果,对算法做了改进,提高了算法的有效性。最后,利用Visual C+ + 6.0实现了一个足球视频内容标注与解析系统,本系统界面简单,具有镜头检测、关键帧提取和内容标注的功能。基本实现了对足球视频内容的解析。更多还原

【Abstract】 With the rapid development of multimedia information processing, computer network, high-speed communications, data networks and the Internet, the digital video applications has been greatly promoted. Such as VOD, Digital TV, Digital Libraries, Video Conferencing, Tele-education, E-commerce and so on. These applications have been accepted and known well by more and more people. Video data with the characteristics of large quantity and a growing number, in the face of a large number of video data, how to search the needed video data(such as the spectacular scenes of a football match)has become an urgent problem.The keyword-based database retrieval is the traditional method. For video data, the method is not valid. Video name enquiries and the play function such as video recorders have been unable to meet people’s needs. Like a book usually have the catalogue and index, these can help people quickly browse and inquire the content, and a video also has the same demand. When the quantity of video resources reached massive level, or the processing speed is needed to near the real-time speed, using traditional video content annotation methods fully by the people will encountered the difficulties that is difficult to overcome.In order to achieve the efficient retrieval for video and other multimedia information, people started to study the "content", which is included in the video, and therefore a new field of study called Content-Based Video Analysis and Retrieval(CBVAR) has been formed. The core of this method is through the computer analysis and understanding for video content, building the structure and semantic indexing, so it’s a great convenience for user to retrieval. This paper mainly studies the content annotation and analytical method of soccer video, through some processes, such as shot detection, key-frame extraction, feature extraction and so on, the shots of soccer video can be classified, and achieving the purpose of content annotation and analysis.Main content of this thesis: Chapter 1 summarizes the conception of CBVAR and the signification of this issue, introduces the researching purpose and main content of this thesis.Chapter 2 introduces the basic knowledge of digital video and fundamental theories of video compression, as well as the MPEG video compression standard, analyzes the MPEG video format and several commonly used compression method. As a common standard of video compression, the MPEG video sequences are more and more widely used, so this paper mainly discuss and research about the MPEG soccer video.Chapter 3 discusses the relevant content of shot detection and key-frame extraction, and introduces a key-frame extraction method which is given in this paper. Shot detection is a most fundamental and important content in CBVAR, and its basic mission is to divide the video sequences into the basic units (the shot) which is relatively independent. The accuracy of shot detection will directly influence the effectiveness of the video retrieval, so the ability of accurately, quickly and effectively detect the shot’s changes, has important implications for the video analysis and retrieval. The types of shot transform can be classified Abrupt Change and Gradual Change, and the approaches to shot detection can be classified as those that operate in non-compressed domain and those in compressed domain. In the non-compressed domain, the basic methods are the pixel comparison method, histogram method and the method based on edge. These methods of shot detection have been very mature, some of them have reached more than 90 percent accuracy, which can satisfy the generally applied requirements.At present the majority of video data are stored in the form of the compression, the research on the shot detection algorithm in compressed domain, is especially important for fast video analysis, segmentation and retrieval. Taking into account the MPEG standard is widely used in video storage, the paper mainly discusses the shot detection methods of MPEG compressed video sequences. In these methods, the way of extracting 8×8 average image (DC image) from the compressed data has been more mature, this approach also applies to other types of video compression standard. Additionally we also extract the information about motion from the compressed data, such as motion vectors, so the shot detection method that based on motion vectors has been formed. In this paper, we implement a method for shot detection using the macroblock (MB) type information, This method can effectively detect the abrupt change and partly gradual change, and the processing speed of this approach is far faster than the non-compressed domain method, so this paper used this method to detect the shots of soccer video.Key frame is a critical image frame which is used to describe the contents of a shot, it usually reflect the main content of a shot, the key-frame extraction play an important role in the CBVAR. This paper introduces several typical key-frame extraction algorithm, and a key-frame extraction method based on statistical frame is given in this paper, the method can be simple and effective extraction 1-2 image as a key frame, which can express the meaning of a shot on the whole.Chapter 4 discusses the key technology of the soccer video analysis, a soccer video shot classification algorithm based on the local feature of the frame image is given in this paper. Soccer is one of the world’s most popular sport, the competition lasted for a long time, and the content of video is rich, this paper mainly studies the method of content annotation and analysis in soccer video. The annotation and analysis of soccer video content can establish an effective content-based indexing feature, and enable users to quickly and easily browse and search the soccer video content, offer the possibility of effective storage and retrieval for the soccer video database, through the realization of soccer content annotation and analysis system, we can provide a clearer structure and the roadmap for soccer video, and achieve the intelligent content-based video-on-demand applications. This paper analyses the key technology of content analysis about soccer video shot, that is the feature extraction, the football video camera analysis of the key technology content, and introduce some commonly used algorithm.According to the characteristics of soccer video, the paper divided the soccer video shot into three types: the spectacular shot, close-up shot and the feature-story shot. A soccer video shot classification algorithm based on the local feature of the frame image is given in this paper. Firstly we divided the key-frame image into blocks in accordance with a certain proportion, then calculated the ratio of green pixels for each block, finally according to the information to achieve the classification of the shot. And users can set the threshold according to their own needs, and obtain a better classification results.Chapter 5 introduces the soccer video content annotation and analysis system which is given in this paper. By Visual C + + 6.0, we implement a soccer video content annotation and analysis system, with the functions such as video play, shot detection and key-frame extraction, and finally we implement the content annotation and analysis of soccer video.Chapter 6 summarizes the main elements of this paper. Based on the current research background and situation, we propose our own views on the follow-up research.更多还原