监控视频异常行为检测算法研究

发布时间：2018-05-07 05:02

本文选题：异常检测 + 多特征提取　；参考：《中国科学技术大学》2017年硕士论文

【摘要】：随着社会的发展与进步,越来越多的社会安全问题也随之出现。如何避免与及时应对突发公共事件作为公共安全的重要议题被学术界及工业界广泛研究,进而产生了多种多样的安防机制及相应的技术。视频监控技术作为被广泛使用的监管技术,在多个方面起到了至关重要的作用。然而目前的通用监控技术仍然存在各种各样的问题。作为计算机视觉的重要范畴,对监控视频内容的分析从发展之初便受到学术界的关注。视频内容的分析主要包括目标检测、目标识别、目标追踪等几大领域。本文主要针对监控视频中的异常行为检测问题,首先对异常的基本概念和基础理论进行了简单介绍,综述和分析了国内外的研究现状和主流算法,随后提出了针对全局和局部异常检测的算法,并分别对算法的具体实施和测试结果进行了相关的分析。本文的研究内容和创新点如下:1.提出了基于多特征提取的局部异常检测算法。传统的异常检测算法通常笼统地对异常进行统一的特征提取,这将造成运动信息的不合理使用甚至丢失,进而造成对异常检测的高误检率。本文通过构建局部异常的特征模型,将局部异常进一步划分为速度异常、形态异常及位置异常,并为每一类异常分别进行特征的提取和检测,并最终得到统一性的异常得分。实验结果证明了这种分类方式是合理而具有优势的。2.提出了基于分裂合并聚类算法的前景目标提取方案。传统的局部异常检测特征提取方案通常将视频序列划分为固定大小规则形状的块(二维)或体(三维),这种划分方式会产生对目标的错误分割及误判率增加等问题。本文通过多次限制性迭代,在每一步迭代中不断分裂不符合样本模型的前景区域,并合并相似性区域,实现对前景的有效分割,提取出单个或多个重叠目标的前景,在此区域内实现特征提取等操作。这一方案对特征区分度的影响较大,因此也是决定整体算法有效性的关键因素。3.改进了 NSH分类算法。原有NSH算法尽管可以实现异常数据的检测,然而其在寻找最优参数时效率较低且鲁棒性较差。本文提出的INSH算法通过在原有二分类算法框架基础上重新定义目标函数和重新设计凸包解决方案,实现比原有算法更优的分类检测结果。4.提出了基于全局动能的全局异常检测算法。传统的全局异常检测算法大多模型复杂、检测效果不尽如人意。本文通过分析大规模运动特点提出了基于全局动能差的全局异常检测算法,算法通过构建全局动能进而计算动能差实现全局异常的检测,实验证明该方法时间成本较低且而行之有效的。
[Abstract]:With the development and progress of society, more and more social security problems appear. How to avoid and respond to unexpected public events as an important issue of public safety has been extensively studied by academia and industry, which has resulted in a variety of security mechanisms and corresponding technologies. As a widely used technology, video surveillance plays an important role in many aspects. However, there are still various problems in the current universal monitoring technology. As an important category of computer vision, the analysis of surveillance video content has been concerned by academia since the beginning of its development. Video content analysis mainly includes target detection, target recognition, target tracking and so on. This paper focuses on the detection of abnormal behavior in surveillance video. Firstly, the basic concepts and theories of anomaly are briefly introduced, and the current research situation and mainstream algorithms at home and abroad are summarized and analyzed. Then an algorithm for global and local anomaly detection is proposed, and the implementation and test results of the algorithm are analyzed respectively. The contents and innovations of this paper are as follows: 1. A local anomaly detection algorithm based on multi-feature extraction is proposed. The traditional anomaly detection algorithm usually carries on the unified feature extraction to the anomaly in general, which will cause the unreasonable use and even the loss of the motion information, and then cause the high error detection rate of the anomaly detection. In this paper, the local anomaly is further divided into velocity anomaly, morphological anomaly and location anomaly by constructing the feature model of the local anomaly, and the feature extraction and detection for each kind of anomaly are carried out respectively, and finally the unified anomaly score is obtained. The experimental results show that this classification method is reasonable and has advantages. 2. A foreground target extraction scheme based on split merge clustering algorithm is proposed. The traditional local anomaly detection feature extraction scheme usually divides the video sequence into blocks (2D) or volume (3D) with fixed size and regular shape. This partitioning method will result in the problem of target segmentation and the increase of misjudgment rate. In this paper, through multiple restrictive iterations, the foreground regions which do not conform to the sample model are split continuously in each iteration, and the similarity regions are merged to realize the effective segmentation of the foreground, and the foreground of single or multiple overlapping targets is extracted. In this region, feature extraction and other operations are implemented. This scheme has a great influence on the classification of features, so it is also a key factor to determine the effectiveness of the whole algorithm. The NSH classification algorithm is improved. Although the original NSH algorithm can detect abnormal data, it is less efficient and less robust in finding the optimal parameters. The INSH algorithm proposed in this paper redefines the objective function and redesigns the convex hull solution on the basis of the original two-classification algorithm framework to achieve a better classification detection result than the original algorithm .4. A global anomaly detection algorithm based on global kinetic energy is proposed. Most of the traditional global anomaly detection algorithms are complex, and the detection results are not satisfactory. In this paper, a global anomaly detection algorithm based on global kinetic energy difference is proposed by analyzing the characteristics of large scale motion. The algorithm realizes global anomaly detection by constructing global kinetic energy and calculating kinetic energy difference. Experiments show that the time cost of this method is low and effective.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】