【作者】 王威；

【导师】 王润生；

【作者基本信息】 国防科学技术大学 ， 信息与通信工程， 2010， 博士

【摘要】 视频处理技术在科学研究和工程应用上有着十分诱人的前景。视频设备的连续工作,产生了大量需要处理的数据。如何对视频事件进行快速而准确的分析是一个值得重点关注的热点问题。论文针对视频事件分析的相关方法进行了深入研究,主要包括:视频序列段落划分、运动目标检测与跟踪和视频语义事件检测。视频序列中的事件主要由两方面原因引起:一是目标的运动或变化情况,二是场景整体的变化情况。因此相应的研究也从两方面展开,分别是基于目标属性约束的视频事件检测和基于复杂条件约束的足球感兴趣事件检测。在视频段落划分方面,提出了一种基于帧间信息的视频段落划分方法,可以从视频序列中检测出时空联合分布上和序列整体平均特性明显不一致的段落。该方法选择颜色变化信息、运动变化信息和运动变化率来描述场景的变化和场景中目标的变化,并对长时间视频序列进行段落划分。方法不需要提取关键帧和运动目标,克服了一般镜头检测方法处理静态背景视频时出现的效果下降问题,并且提高了处理效率。在运动目标检测与跟踪方面,提出了一个既适用于静态背景又适用于动态背景的目标检测和跟踪处理框架。框架由三个部分组成:基于连续帧差的背景类型确定,采用粒子滤波方法的目标跟踪,以及采用适合具体背景类型策略的目标精确检测。在静态背景中,提出了一种改进的自上而下的局部层次化混合高斯模型算法(LHGMM)进行目标精确检测,由于方法在局部区域内进行相应计算,可以提高准确性和处理效率。在动态背景中,提出了一种自适应水平集(ALS)方法进行目标轮廓精确检测。方法能够自动确定零水平集并在特定区域内进行曲线演化,使目标轮廓检测更加精确。在基于目标属性约束的视频事件检测方面,首先提出一种基于模糊粗糙集的属性选择方法。该方法能够结合具体应用背景,合理选择视频处理过程需要的特征,并建立事件的时空联合描述。然后研究了三种具体视频场景下的视频事件分析方法:(1)提出一种基于区域特征的过路行人异常行为检测方法,利用背景区域分割信息和目标区域变化信息检测过路行人的异常行为,可以满足实时性处理要求。(2)提出一种基于轮廓特征的人体姿态分析方法,利用轮廓特征的周期特性对人体姿态进行分类。(3)提出一种基于运动特征的交通路口视频事件检测方法,通过对运动轨迹的自动学习分析运动轨迹的区域和方向,并检测异常行为事件,实现监控视频的在线处理。在基于复杂条件约束的足球事件检测方面,首先提出一种基于层次化分类树模型的视频片段分类方法。该方法仅利用简单的低层特征,就能够对提取的视频片段实现迅速而有效的分类。在片段分类的基础上,提出一种基于时间结构信息的足球感兴趣事件检测方法。方法利用足球视频事件特定的时间结构信息与运动等低层特征相结合检测预先设定的感兴趣足球事件。由于检测方法直接利用场景条件约束,避免了难以实现的大量运动目标的检测与跟踪,提高了处理质量与效率。更多还原

【Abstract】 The technology of video processing has promising future in science research and engineering application. Due to the continuous working of the equipments, a lot of video sequences are produced and need dealing with. How to analyze the video events rapidly and exactly is a hotspot which captures much attention. The thesis focuses on the analysis methods of the video events. The research includes the following aspects: division of the video sequence, moving targets detection and the tracking, and semantic events detection. Generally there are two kinds of reasons which cause the events: the motion or the change of the targets, and the change of the whole scene. So the corresponding research also outspread in two sides, namely events detection based on the restrictions of the targets’characters and events detection based on the restrictions of complex conditions in soccer games.In division of the video sequences, a novel inter-frame-information-based approach is proposed to detect the sections whose spatio-temporal distribution is obviously different from the average distribution of the whole video sequence. In this approach, the color information, motion information and moving ratio information are used to describe both the change of the scene and the changes of the targets in the scene, and divide the long video sequence into sections. This approach does not need to extract the key frames, and overcomes the ineffective flaw of the shot boundary detection methods when the background is stable. The processing efficiency is also improved at the same time.In the moving targets detection and tracking, a novel framework is proposed to process the video when the background is either stable or dynamic. It consists of three aspects: background type recognition based on difference information of the adjacent frames, targets tracking with the color-based particle filter, and objects precise detection by using the strategies which is suitable for special background types. In stable background, a top-to-bottom local hierarchical GMM (LHGMM) is proposed to detect the targets accurately. This approach only detects the targets in local regions and can improve both the veracity and the efficiency. On the other hand, when the background is dynamic, an adaptive level set (ALS) method is proposed to get the precise external contour of the targets. This method can set the zero level set automatically and evolution the level set curve in given areas. So this method can get the accurate external contour easily.In the events detection based on the restrictions of the targets’characters, the characters selection approach based on fuzzy-rough techniques is firstly proposed. This approach can be used to select the characters combined with the application conditions, and describe the events by using the spatio-temporal information. Then, the events analysis is carried out in three special aspects: (1) A region-based abnormal behaviors detection approach of the road-across pedestrian is proposed to detect the abnormality by using both the region-based segmentation information and the change information of the targets. This approach can satisfy the on-line video process requirement. (2) A contour-based approach is proposed to analyze the human poses. This approach uses the periodic characters of the moving human to categorize the human poses. (3) A motion-based detection approach of the video events in the traffic crossing scene is proposed to detect the abnormal behaviors by using the moving trajectories. With the self-acting study on moving trajectories, the approach can get the regions and the moving direction of the moving trajectories, and detect the abnormal behaviors. So the surveillance systems can work on-line.In the events detection based on the restrictions of the complex conditions in soccer games, a hierarchical classification tree is proposed to classify the video clips rapidly and effectively only by using simple low-level characters. Then based on the clips classification, a temporal structures-based approach is proposed. This approach can detect the prior-defined exciting events by using both fixed temporal structure of clips and the low-level characters, such as motion vector and so on. Neither the tracking of the targets nor the prior training is necessary because this approach uses the restrictions of the scene directly, and this can improve both the processing quality and the efficiency.更多还原