计算机工程

• 图形图像处理 • 上一篇    下一篇

利用综合光流直方图的人群异常行为检测

熊饶饶 1,胡学敏 1,陈龙 2,周慧子 1   

  1. (1.湖北大学 计算机与信息工程学院,武汉 430062; 2.中山大学 数据科学与计算机学院,广州 510006)
  • 收稿日期:2016-08-29 出版日期:2017-10-15 发布日期:2017-10-15
  • 作者简介:熊饶饶(1995—),男,本科生,主研方向为图像处理、智能视频分析;胡学敏(通信作者)、陈龙,讲师、博士;周慧子,本科生。
  • 基金项目:

    国家自然科学基金青年科学基金(41401525);湖北省大学生创新创业训练计划项目(201510512041)。

Abnormal Crowd Behavior Detection Using Synthesized Optical Flow Histogram

XIONG Raorao 1,HU Xuemin 1,CHEN Long 2,ZHOU Huizi 1   

  1. (1.School of Computer Science and Information Engineering,Hubei University,Wuhan 430062,China; 2.School of Data and Computer Science,Sun Yat-Sen University,Guangzhou 510006,China)
  • Received:2016-08-29 Online:2017-10-15 Published:2017-10-15

摘要:

针对公共区域下的智能视频监控问题,提出一种新的从视频中检测人群异常行为的方法。利用混合高斯模型提取视频中的人群运动前景,在运动前景区域内使用等间距抽样法提取特征点。在人群特征提取阶段,给出光流特征提取方法,通过Lucas-Kanade法计算特征点的光流场,并统计计算全局特征点的光流方向直方图、光流大小直方图和光流加速度直方图,将融合3种直方图的综合光流直方图作为人群特征,使用支持向量机对特征数据进行训练和预测,判断人群中是否存在异常行为。实验结果表明,与基于社会力模型和纯光流方向直方图的方法相比,该方法能够有效、实时地检测人群中的异常行为,在UMN数据集中的检测率达到97%以上。

关键词: 异常行为检测, Lucas-Kanade光流, 特征点提取, 运动矢量, 支持向量机

Abstract:

To deal with the issue of intelligent video surveillance in public places,a novel method detecting abnormal crowd behavior from videos is proposed.The moving foreground is extracted from the video by using the Gaussian Mixture Model(GMM).Feature points are extracted from foreground regions by an equidistant sampling method.In the stage of crowd feature extraction,an optical flow feature extraction method is presented,where the Lucas-Kanade method is used to calculate the optical flow field.Crowd features are constructed by synthesizing three kinds of histograms including orientation,magnitude and acceleration of the optical flow.The Support Vector Machine(SVM) is applied to train and predict the feature data from the total histogram.Experimental results show that the proposed method can effectively detect abnormal crowd behaviors in real time compared with the methods based on social force model and pure histogram of optical flow.The detection rate in the UMN dataset is greater than 97%.

Key words: abnormal behavior detection, Lucas-Kanade optical flow, feature point extraction, motion vector, Support Vector Machine(SVM)

中图分类号: