• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学

• 论文 • 上一篇    下一篇

基于隐马尔可夫模型的视频异常场景检测

李娟1,张冰怡1,冯志勇1,徐超2,张铮3   

  1. (1.天津大学计算机科学与技术学院,天津 300350;2.天津大学软件学院,天津 300350;
    3.天津工业大学计算机科学与软件学院,天津 300387)
  • 收稿日期:2016-01-06 修回日期:2016-03-29 出版日期:2017-07-25 发布日期:2017-07-25
  • 基金资助:

    国家自然科学基金(2008CB987656)

Anomaly detection based on hidden Markov model in videos

LI Juan1,ZHANG Bing-yi1,FENG Zhi-yong1,XU Chao2,ZHANG Zheng3   

  1. (1.School of Computer Science and Technology,Tianjin University,Tianjin 300350;
    2.School of Computer Software,Tianjin University,Tianjin 300350;
    3.School of Computer Science & Software Engineering,Tianjin Polytechnic University,Tianjin 300387,China)
  • Received:2016-01-06 Revised:2016-03-29 Online:2017-07-25 Published:2017-07-25

摘要:

视频技术的广泛应用带来海量的视频数据,仅依靠人力对监控视频中的异常进行检测是不太可能的。异常行为的自动化检测在公共安全等领域的地位极其重要。提出一种综合考虑目标特性和时空上下文的异常检测方法,该方法利用光流纹理图描述移动物体的刚性特征,建立基于隐马尔可夫模型HMM的时间上下文异常检测模型。在此基础上,提取异常目标的Radon特征,以支持向量机SVM的异常预分类结果为基础,通过HMM建立异常场景的空间上下文分类模型。该模型在公共数据集UCSD PED2上进行了实验验证,结果表明,本算法不仅在异常检测方面优于已有算法,而且还能给出异常分类。

关键词: 刚体, 隐马尔可夫模型, 人群场景, 异常检测

Abstract:

Widely used video technology brings a large amount of video data, so it is impossible to detect abnormalities in surveillance video relying on human operators only. Automated abnormal events detection is extremely important for public safety. We propose an abnormal events detection approach with comprehensive consideration of target features and temporal-spatial context. The method exploits the texture of optical flows to describe the rigidity of moving objects. Then, we establish an abnormal events detection model of temporal context based on the hidden Markov model (HMM). Afterwards, radon features of abnormal events are extracted. We also establish the classification model of spatial context by using the HMM based on pre-classification results obtained by the support vector machine (SVM). Experimental results on public dataset—UCSD PED2 show that the performance of our method outperforms the existing algorithms in abnormal events detection and localization. Furthermore, our approach can classify abnormal events.
 

Key words: rigid, hidden Markov model, crowd scenes, abnormal events detection