【作者】 胡滨；

【导师】 王秀坤；

【作者基本信息】 大连理工大学 ， 软件工程， 2006， 硕士

【摘要】 随着体育视频的不断普及,从大量的视频数据中手动的寻找关键片断是一件令人厌烦和乏味的工作,对关键画面自动检测系统的需求也越来越强烈。 本文给出了一种有效的足球视频结构分析系统,该系统可自动分析足球视频中的关键事件,根据电影特征和对象特征生成视频概要。系统主要包括镜头边界检测、镜头分类、慢动作回放镜头检测和生成视频概要四个部分。由于足球视频的特殊性,在镜头边界检测中采用分层检测的算法,第一层为像素点对的比较和直方图的比较,第二层为对象分割和对象跟踪。在镜头边界检测完成之后进行镜头分类,一种分类方式是按照足球场区域将足球场划分为9部分。另一种分类方式是根据足球视频背景的特殊性,即足球场草坪是以绿色为主色调,按照这个主要特征将镜头分成远镜头、中镜头、人物特写或场外镜头三类。在镜头分类中,本文采用了黄金分割算法和最大非草色矩形算法来划分中、远镜头。在分析进球事件的基础上,进一步分析慢动作回放镜头的特点和镜头间的关联规则来寻找关键片断,进行视频概要的抽取。 本系统能输出足球比赛中慢动作回放的镜头和足球比赛中的关键事件。前一类概要仅基于电影特征和底层特征快速生成的,而后一类概要包含了高层语义特征。根据电影特征和底层特征就足够获得特定的关键事件时,就没有必要再计算基于对象的特征。当结果需要提高准确率时可以花费更多的代价来计算基于对象的特征。抽取的足球视频概要能用来进行基于内容的视频检索。更多还原

【Abstract】 As digital sports video data become more and more pervasive, finding the clips of highlights manually in large amount of video data is a boring and tedious task, automatic highlight detection system is particularly demanding.In this paper, it presents an automatic and effective framework for highlight events detection based on cinematic and object features to get video summarization. The proposed framework includes 4 parts which are shot boundary detection, shot classification, slow-motion replay shot detection and video summarization. The shot boundary detection employs the multi-filter structure that combines traditional pixel level comparison and histogram comparison with object segmentation and object tracking for the particularity in soccer video. After the shot boundary detection, there are two different methods for the shot classification. The shots can be classified by the soccer field region, and the soccer field can be divided into 9 parts. The shots can also be classified by the background of the soccer video which the domain color is the tone of green. The soccer shots are classified into three classes which are long shots, medium shots, out of field or close-up shots by using this feature. It presents gold section and finding the max ungreen pixel rectangle algorithms to partition long shot and medium shot in shot classification. Then, the highlight event uses the slow-motion replay and association rule to locate the highlight shot and the new template to identify the highlight event and extract summarization on the basis of goal event.The system can output two types of summaries which are all slow-motion segments in a game and all highlight events in a game. The first type of summaries is based on cinematic and low-level features only for speedy processing, while the summaries of the second type contain higher-level semantics. It is efficient in the sense that there is no need to compute object-based features when cinematic and low-level features are sufficient for the detection of certain events. It is effective in the sense that the framework can also employ object-based features when needs to increase accuracy. The summarization that is extracted from soccer video can be used to implement content-based video retrieval.更多还原