基于监控视频的人群异常事件检测

发布时间：2018-05-18 18:20

本文选题：异常事件检测 + 排斥力　；参考：《郑州大学》2017年硕士论文

【摘要】：近年来,随着越来越多人群踩踏、恐怖袭击以及其他公共安全事件的发生,作为群体行为监控、分析、预警基础的人群异常事件检测已经成为智能监控领域中亟需解决的问题之一。然而,由于人群运动复杂多变,要列出人群中所有可能出现的异常事件几乎是不可能的,因此人群异常事件的检测并不是一个典型的分类问题。为此本文采用先从只包含正常事件的人群视频中训练出一个特征集,再通过计算待检测视频中人群特征与特征集的偏离程度来判断人群中是否出现异常这一技术路线。具体地,本文将人群异常事件的检测分为两步:一是人群视频特征提取,即事件表示;二是模型的训练更新和人群异常事件检测。对于人群视频的特征提取,本文采用了两种不同方法。一是基于排斥力的特征提取。为避免在拥挤场景中跟踪每个具体个体带来的问题,本文使用一种粒子的平流运动来模拟人群的运动,并将通过排斥力模型获取的能够准确反映出人群运动的力流矩阵视为提取的特征。二是基于卷积神经网络的特征提取。虽然第一种方法提取的特征在多种类型场景的异常检测中都取得了良好的效果,但是它对于排斥力变化不明显的场景敏感度不高。因此我们采用一种改进的卷积神经网络模型,并结合滑动窗口和PCA以提取人群视频的时空特征。对于模型的训练和异常事件检测,本文采用了一个基于稀疏编码的分组词典框架以改善单一模式词典带来的计算量大且难以维护的问题。在异常事件检测时,本文依次使用每个词典对待检测单词进行稀疏重构。一旦待检测单词能被某个词典稀疏表示,那么就认为它所表示的事件是正常的;若所有的词典都不能稀疏表示它,那么就认为是异常的。为解决随着新视频的不断增加和视频场景的动态变化而导致的词典组表示能力退化或概念漂移问题,本文基于稀疏重构和一组单词池提出了一个无监督的词典组局部和全局在线更新算法。最后,本文在公开数据集UMN、UCSD以及Web数据集上与其他方法进行了对比分析。实验表明本文的两种方法都提高了人群异常事件检测的准确度和效率,并且基于卷积神经网络特征提取的方法改善了基于排斥力的特征提取方法对于排斥力变化不明显场景敏感度不高的问题。
[Abstract]:In recent years, as more and more people stampede, terrorist attacks and other public safety incidents, as group behavior monitoring, analysis, Detection of abnormal events based on early warning has become one of the most urgent problems in intelligent monitoring field. However, due to the complexity of crowd movement, it is almost impossible to list all possible abnormal events in the crowd, so the detection of abnormal events is not a typical classification problem. In this paper, we first train a feature set from a crowd video containing only normal events, and then calculate the deviation between the crowd feature and the feature set to determine whether there is an anomaly in the crowd. Specifically, the detection of crowd abnormal events is divided into two steps: one is crowd video feature extraction, that is, event representation; the other is model training update and crowd abnormal event detection. For crowd video feature extraction, this paper adopts two different methods. One is the feature extraction based on repulsive force. To avoid the problem of tracking each individual in a crowded scenario, a particle advection motion is used to simulate the movement of a population. The force flow matrix obtained by repulsive force model can accurately reflect the movement of people is regarded as the feature of extraction. Second, feature extraction based on convolution neural network. Although the feature extracted by the first method has achieved good results in anomaly detection of many kinds of scenes, it is not sensitive to scenes where the repulsive force is not obvious. So we adopt an improved convolution neural network model and combine sliding window and PCA to extract temporal and spatial features of crowd video. For model training and anomaly event detection, a group dictionary framework based on sparse coding is used to improve the computational complexity and difficult maintenance of single pattern dictionary. In anomaly event detection, each dictionary is used for sparse reconstruction of detected words. Once a word to be detected can be represented sparsely by a dictionary, it is assumed that the event it represents is normal; if all dictionaries cannot represent it sparsely, then it is considered abnormal. In order to solve the problem of loss of ability or concept drift of dictionaries caused by the increasing number of new videos and the dynamic changes of video scenes, Based on sparse reconstruction and a set of word pools, an unsupervised online updating algorithm for local and global dictionary groups is proposed in this paper. Finally, this paper compares and analyzes the UMN / UCSD and Web datasets with other methods. The experiments show that the two methods in this paper improve the accuracy and efficiency of the detection of crowd abnormal events. And the feature extraction method based on convolution neural network improves the problem that the feature extraction method based on repulsive force is not sensitive to the change of repulsion force.
【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】