基于光流共生矩阵的人群行为异常检测-《郑州大学学报(工学版)》

文章信息/Info

Title:: Detection of Human Behavior Anomaly Based on the Optical Flow Co-occurrence Matrix

Author(s):: Zeng Qingshan; Song Qingxiang; Fan Mingli; School of Electrical Engineering, Zhengzhou University, Zhengzhou, Henan 450001

Keywords:: crowd behavior anomaly detection; optical flow; optical flow co-occurrence matrix; support vector machine

摘要:: 针对传统基于图像纹理特征进行人群行为异常检测算法倾向于描述人群图像纹理的变化而非人群运动行为的实际情况，导致检测性能较差的问题，提出一种反映人群运动行为真实情况的特征提取方法。首先通过Lucas-Kanade光流算法提取人群视频光流信息，并建立光流幅值共生矩阵与光流方向共生矩阵，然后通过共生矩阵提取角二阶距、对比度、熵、相似度等特性，并将其与光流幅值均值合并组成特征向量训练支持向量机，最后判断人群行为和是否有异常。仿真结果表明，本文的特征提取方法更加深化地处理了光流法提供的人群运动信息，具有较好的人群异常行为识别性能。

Abstract:: The traditional anomaly detection algorithm for human beharior that based on image texture featurestended to describe the change of human image texture ranther than the actual situationof human motion behavior. Its detection porformance was not so good. In this paper, a method of feature extraxtion was proposed to reflect the real situation of human motion behavior. Firstly, the optical flow information was extracted by Lucas-Kanade optical flow algorithm, and the co-occurrence matrix and optical flow direction co-occurrence matrix were established. Then, the characteristics of two order distance,contrast, entropy and similarity are extraced by the co-occurrence matrix, and then combined them with the mean value of optical flow to form a feature vector to train the support vector machine(SVM). Finally, this algorithm was used to determine whether the crowd had abnormal behavior. The simulation results showed that the feature extraction method in this paper had more in depth processing of the croed motion information provided by the optical flow method.Compared with the mainstream algorithm, it has a better recognition performance.