基于视频的人群行为异常检测

发布时间：2018-05-22 07:29

本文选题：人群行为异常检测 + 修正社会力模型　；参考：《郑州大学》2017年硕士论文

【摘要】：随着我国经济社会的迅速发展,城市化进程的不断加快,城镇人口密度的不断上升,公共场所安全事故频繁发生,传统的视频监控技术已无法满足社会发展进步的需要,迫切需要大力发展智能视频监控技术。人群行为异常检测作为当今智能视频监控研究领域的热点问题之一,主要包括对视频图像序列进行运动目标检测、特征提取、跟踪、识别与行为分析等内容。因涉及到计算机视觉、图像视频处理与人工智能领域众多核心问题,对人群行为异常检测算法的研究也是一个具有很高挑战难度的问题。本文主要针对公共场所人群骚乱、逃散等人群行为异常状态进行特征提取与检测,分别从不同角度提出了基于修正社会力模型的人群行为异常检测算法与基于光流共生矩阵的人群行为异常检测算法,研究内容涉及人群运动信息提取、特征点检测、行人社会力建模与计算、光流共生矩阵建立与特征提取、特征分类等。主要研究工作如下:1)针对经典社会力模型的不足,通过引入行人之间的相对速度与相对位置对行人施加的不同影响因素对其进行了修正;通过研究现有基于社会力模型的人群行为异常检测算法,针对其利用网格粒子采样方法进行行人社会力计算方法中存在的缺陷,提出了一种利用行人Harris角点进行行人社会力计算的改进算法,首先利用Lucas-Kanade光流算法提取视频行人运动信息,然后利用Harris角点检测算法对行人进行角点检测,根据修正后的社会力模型对行人角点进行社会力计算,最终基于社会力进行人群行为状态特征提取。2)通过对基于动态纹理特征的人群行为异常检测算法的学习与研究,发现动态纹理特征本质上是一种图像灰度纹理的特征描述子,并不能如实的反映人群的运动信息。本文根据不同人群行为状态下行人运动信息空间分布特点,提出通过建立光流共生矩阵的方法提取人群运动信息分布特征,进行人群行为状态的特征提取。3)基于支持向量机分类器分别阐述了基于修正社会力模型的人群行为异常检测算法与基于光流共生矩阵的人群行为异常检测算法的实验流程。选择不同的视频数据集对算法性能进行了实验验证,并将本文算法实验结果与参考文献中的算法实验结果进行对比,证明了本文算法在人群行为异常检测问题上性能的优越性。
[Abstract]:With the rapid development of our country's economy and society, the accelerating process of urbanization, the increasing of urban population density and the frequent occurrence of safety accidents in public places, the traditional video surveillance technology has been unable to meet the needs of social development and progress. It is urgent to develop intelligent video surveillance technology. Crowd behavior anomaly detection is one of the hot issues in the field of intelligent video surveillance. It mainly includes moving target detection, feature extraction, tracking, recognition and behavior analysis of video image sequences. Because of the many core problems in computer vision, image and video processing and artificial intelligence, the research of crowd behavior anomaly detection algorithm is also a very challenging problem. In this paper, the characteristics of abnormal behavior of people in public places, such as riots and desertions, are extracted and detected. From different angles, this paper proposes a modified social force model based crowd behavior anomaly detection algorithm and an optical flow symbiosis matrix based crowd behavior anomaly detection algorithm. The research content involves crowd motion information extraction, feature point detection. Pedestrian social force modeling and calculation, optical flow symbiosis matrix building and feature extraction, feature classification and so on. The main research work is as follows: (1) aiming at the deficiency of the classical social force model, this paper modifies it by introducing the different factors of the relative velocity and relative position of the pedestrian. Based on the existing social force model based on the crowd behavior anomaly detection algorithm, aiming at its use of mesh particle sampling method for pedestrian social force calculation method shortcomings, An improved algorithm for calculating pedestrian social force using pedestrian Harris corner is proposed. Firstly, the video pedestrian motion information is extracted by Lucas-Kanade optical flow algorithm, and then the corner detection algorithm of Harris is used to detect pedestrian corner. According to the modified social force model to calculate the social force of pedestrian corner, finally based on social force to extract crowd behavior state feature. 2) through the study and research of crowd behavior anomaly detection algorithm based on dynamic texture feature. It is found that the dynamic texture feature is essentially a feature descriptor of grayscale texture of an image, and it can not reflect the movement information of the crowd truthfully. According to the spatial distribution characteristics of pedestrian movement information under different crowd behavior, this paper proposes a method to extract the distribution characteristics of pedestrian movement information by establishing optical flow symbiosis matrix. (3) based on support vector machine classifier, the experimental flow of abnormal detection algorithm of crowd behavior based on modified social force model and optical flow co-occurrence matrix is described respectively. Different video data sets are selected to verify the performance of the algorithm, and the experimental results of this paper are compared with the experimental results in the references, which proves the superiority of the proposed algorithm in crowd behavior detection.
【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TN948.6

【相似文献】