基于Vibe改进算法的中高密度人群异常检测方法研究

发布时间：2018-04-30 08:17

本文选题：前景检测 + 人群运动特征　；参考：《西安理工大学》2017年硕士论文

【摘要】：人们的公共安全意识在不断增强,众多公共场所的人群管理具有重要的现实意义。目前人群密度估计与人群异常行为检测技术是人群监控管理的两个重要方面,具有极大的实用价值。针对人群突然向四处逃跑、群殴等人群异常事件,本文结合人群运动特征和人群密度特征来检测人群异常行为,研究内容涉及前景提取算法、特征点提取与跟踪算法、人群密度特征提取方法、人群运动特征提取方法以及人群异常检测算法。本文主要研究工作如下:(1)研究分析常用运动目标检测算法,在此基础上本文设计实现了一种Vibe改进算法,改进算法能改善Vibe算法中产生的鬼影及阴影问题,并且在光照突变下也能进行有效检测,测试结果表明了本文改进算法可以较准确地检测出场景中的运动目标。(2)在提取到前景运动目标之后,本文提出使用运动特征变化率表征人群的运动信息,首先通过金字塔LK光流法跟踪前景图像的特征点计算出运动目标的运动向量,并统计视频帧块区间内的特征点来表征人群的分布状态,结合运动向量和分布状态形成人群运动特征变化率描述人群运动信息。随后通过前景图像像素数对人群密度视频图像分类,若为稀疏人群图像则提取前景特征点数、前景像素总面积、边缘特征和周长面积比特征描述人群密度信息,若为密集人群图像则提取前景图的局部二值模式灰度共生矩阵特征和灰度图的局部二值模式灰度共生矩阵特征描述人群密度特征。(3)研究分析现有人群异常检测方法,在此基础上本文将人群运动特征和人群密度特征结合检测人群异常行为。通过采用标准视频数据集UMN数据集对采取人群运动特征、人群运动和方向特征、人群运动特征变化率的方法进行实验对比,实验验证了本文方法在准确率和实时性上较好的表现。
[Abstract]:People's public safety consciousness is increasing constantly, the crowd management of numerous public places has important practical significance. At present, crowd density estimation and abnormal behavior detection are two important aspects of crowd monitoring and management, and have great practical value. In view of the abnormal events of the crowd, such as the sudden flight of the crowd and the fight, this paper combines the characteristics of the crowd movement and the density of the crowd to detect the abnormal behavior of the crowd. The research involves the foreground extraction algorithm, the feature point extraction and tracking algorithm. Population density feature extraction method, crowd motion feature extraction method and crowd anomaly detection algorithm. The main work of this paper is as follows: (1) Research and analysis of commonly used moving target detection algorithms. On this basis, an improved Vibe algorithm is designed and implemented. The improved algorithm can improve the ghost and shadow problems in the Vibe algorithm. The test results show that the improved algorithm can accurately detect the moving target in the scene. In this paper, the motion information of the population is represented by the change rate of the motion feature. Firstly, the motion vector of the moving object is calculated by tracking the feature points of the foreground image by the pyramid LK optical flow method. The feature points in the video frame block interval are counted to represent the distribution state of the population, and the population motion information is described by combining the motion vector and the distribution state to form the change rate of the crowd motion characteristics. Then the population density video image is classified by the pixel number of foreground image, the foreground feature points are extracted if the image is sparse, the total area of foreground pixel, edge feature and circumference area ratio feature to describe the population density information. If it is a dense crowd image, it extracts the local binary pattern gray level co-occurrence matrix feature of foreground map and the local binary pattern gray level co-occurrence matrix feature of gray scale map to describe the population density characteristic. On this basis, the characteristics of population movement and population density are combined to detect the abnormal behavior of the population. By using the standard video data set UMN data set, the methods of adopting crowd motion feature, crowd movement and direction feature, and the change rate of crowd motion feature are compared experimentally. Experimental results show that the proposed method performs well in accuracy and real-time performance.
【学位授予单位】：西安理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：D63;TP391.41

【参考文献】