视频监控下实时异常行为检测研究

发布时间：2018-01-08 10:23

本文关键词：视频监控下实时异常行为检测研究　出处：《南京邮电大学》2016年硕士论文　论文类型：学位论文

【摘要】：随着计算机和互联网技术的不断发展和社会各个领域对安全需求的不断增长,智能监控技术得到了飞速发展。在机场和火车站的安检、交通管控、国防安全等方面都在应用智能监控技术。传统的视频分析主要依靠人,这种监控系统无法适应现在的海量的视频数据,无法保证监控系统的高效和可靠。高效、可靠的智能监控系统应运而生,能够提高视频监控系统的智能化。在人群中的异常行为检测方法中,本文首先对视频集合进行预处理从而得到可以直接进行处理的视频序列,之后提取全局的目标特征(如时空特征、运动矢量特征、运动矢量直方图、网格尺度通量特征),然后根据这些特征对视频进行异常行为分析,即进行以下三步处理:根据提取的特征能量信息进行异常行为分析;若不存在异常,则根据提取特征的色散性进行异常行为分析;若不存在异常,则根据提取的特征之间的拉格朗日粒子动力学进行分析,若依然不存在异常则进行局部异常检测。在检测过程中,网格尺度特征不仅可以用于全局异常行为检测,而且可以用于对不存在全局异常行为的视频进行局部异常检测。接下来进行局部异常行为的检测,首先对正常活动进行图像块提取,对提取的图像块的关键区域进行建模并计算区域之间的关联性,从而得到正常活动模型,然后对测试视频进行图像块的提取和图像块的优化,从而判断是否存在局部异常行为,若存在异常行为则在图像帧上进行标注。在UMN数据集、UCSD数据集和Subway数据集上使用本文方法进行实验,从实验结果中可以得出本文的方法可以在存在光照、阴影等干扰的情况下,依然能够对视频中的异常行为进行检测,并且较其他几种方法在性能上有所提高。
[Abstract]:With the continuous development of computer and Internet technology and the security requirements of each field of the society, intelligent monitoring technology has been rapid development. At the airport and train station security, traffic control, national security and other aspects in the application of intelligent monitoring technology. Video analysis rely mainly on the traditional people, this can not adapt to the video monitoring system now the massive data, can not guarantee the efficient and reliable monitoring system. The intelligent monitoring system came into being efficient, reliable, can improve the intelligent video monitoring system. The abnormal behavior in the crowd detection method, this paper firstly preprocess the video sequence to get can be processed directly to the video collection target features the overall extraction (such as temporal and spatial features, the feature of motion vector, motion vector histogram, grid scale fluxes), then according to the characteristics of Analysis of abnormal behavior of the video, the following three steps: according to the analysis of abnormal behavior characteristics of energy information extraction; if there is abnormal, abnormal behavior analysis based on dispersion feature extraction; if there is no exception, according to the Lagrange particle dynamics between the extracted features are analyzed, if still there is no exception is the local anomaly detection. In the detection process, the grid scale features can be used not only for global anomaly detection, but also can be used for abnormal behavior there is no global video local anomaly detection. Then detection of abnormal behavior, the first image block extraction of normal activities, and the model correlation calculation between regions the extraction of image blocks of key areas, so as to obtain the normal activity model, then the image blocks were extracted and the test video Optimization of block image, and to determine whether there is abnormal behavior, if there is abnormal behavior is marked in the image frame. In the UMN data set, using the UCSD data set and Subway data set on the experiment method from the experimental results can be obtained in this paper can be illuminated in the presence of shadows, disturbances next, still able to detect abnormal behaviors in video, and increased compared with the other method in performance.

【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TN948.6;TP391.41

【相似文献】