基于隐马尔科夫模型的人群异常场景检测

发布时间：2018-04-28 13:33

本文选题：人群场景 + 异常检测　；参考：《天津大学》2016年硕士论文

【摘要】：监控设备的普及催生了大量的监控数据,使得对监控视频中的异常进行人工检测变得非常困难。为了减轻人力资源和经济负担,同时提高异常检测的准确率,人们不断寻求对视频当中的异常进行自动检测的方法。人群密集的场景更是事故的多发场景,因此,如何对人群异常场景进行自动检测尤为重要。本文提出了一种基于隐马尔科夫模型(HMM)的方法来进行人群中的异常检测和分类。本文将人群中的异常检测定义为密集人群中出现异常的目标:机动车、自行车、滑板;另外还有人群的群体异常行为:人群突然的逃散。本文的目的就是要将这些异常目标和群体异常检测出来,并将异常目标分类。隐马尔科夫模型(HMM)可以利用变量的时空上下文关系为变量建立模型,本文选用隐马尔科夫模型(HMM)进行监控视频中的人群异常检测也正是基于隐马尔科夫模型(HMM)的这种特点。在异常检测阶段,首先利用光流纹理描述运动物体的刚性特征,利用这一特征获得异常的预检测结果,在此基础上利用隐马尔科夫模型(HMM)建立时间上下文的异常检测模型,然后利用Viterbi算法解码获得最优隐状态序列,这个隐状态序列就是异常检测的结果。在获得异常检测结果的同时获得异常目标的所在位置。在异常分类阶段,将异常目标的Radon特征与SVM分类器结合,得到异常目标预分类的结果,并利用异常目标的时间上下文关系建立基于隐马尔科夫模型(HMM)分类模型,解码获得异常分类的结果。为了验证本文算法的有效性,我们在UCSD PED2和UMN数据库中进行实验。实验分为两个阶段,首先对人群中的异常目标和群体异常行为进行检测,然后将检测出的正确异常目标进行分类。最终实验结果表明,本文算法可以准确的对异常进行检测、定位,并对异常进行有效分类。
[Abstract]:The popularity of monitoring equipment has given birth to a large amount of monitoring data, which makes it very difficult to detect anomalies in surveillance video manually. In order to reduce the burden of human resources and economy, and improve the accuracy of anomaly detection, people are constantly looking for automatic detection of anomalies in video. Crowd-intensive scene is the frequent scene of accidents, so it is very important to detect the abnormal scene automatically. In this paper, a method based on Hidden Markov Model (hmm) is proposed to detect and classify abnormal population. In this paper, the abnormal detection in the crowd is defined as the abnormal target in the dense crowd: motor vehicle, bicycle, skateboard, and the abnormal behavior of the crowd: the sudden dispersal of the crowd. The purpose of this paper is to detect and classify these abnormal targets and population anomalies. The Hidden Markov Model (HMMM) can be used to establish the model by using the temporal and spatial context of the variable. In this paper, the Hidden Markov Model (HMMM) is used to detect the abnormal crowd in the surveillance video, which is based on the Hidden Markov Model (HMMM). In the phase of anomaly detection, the rigid feature of moving object is described by optical flow texture, and the pre-detection result of anomaly is obtained by using this feature. Based on this, an anomaly detection model of time context is established by using Hidden Markov Model (HMMM). Then Viterbi algorithm is used to decode the optimal hidden state sequence, which is the result of anomaly detection. The location of the abnormal target is obtained at the same time the result of anomaly detection is obtained. In the phase of anomaly classification, the Radon features of abnormal objects are combined with the SVM classifier to obtain the results of pre-classification of abnormal objects, and the hmm) classification model based on Hidden Markov Model (hmm) is established by using the temporal context of abnormal objects. Decode the result of abnormal classification. In order to verify the validity of this algorithm, we have carried out experiments in UCSD PED2 and UMN databases. The experiment is divided into two stages. First, the abnormal target and the abnormal behavior of the population are detected, and then the correct abnormal targets are classified. Finally, the experimental results show that the algorithm can accurately detect, locate and classify the anomalies.
【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;TN948.6

【参考文献】