SIA OpenIR  > 光电信息技术研究室
面向智能监控的人类异常行为检测方法研究
Alternative TitleResearch on Human Abnormal Behavior Detection for Intelligent Surveillance
周培培1,2
Department光电信息技术研究室
Thesis Advisor丁庆海 ; 罗海波
ClassificationTP277
Keyword计算机视觉 视频分析 人群异常检测 暴力异常检测 人数异常检测
Call NumberTP277/Z75/2018
Pages137页
Degree Discipline模式识别与智能系统
Degree Name博士
2018-05-18
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract本课题首先对异常行为进行表征,把异常行为分为非特定异常行为和特定异常行为两类,其中,非特定异常行为主要研究了人群异常行为,特定异常行为选取暴力异常和人数估计进行研究,并针对不同的异常检测算法作了综述。针对不同的异常检测任务和遇到的挑战,本课题给出了不同的解决方案。论文的主要贡献如下:(1) 提出了一种基于运动轨迹特征的人群异常检测算法。该算法以跟踪算法获取的轨迹点为研究对象,首先,根据轨迹点的速度、角度等属性把轨迹表示成不同线段;然后采用DBSCAN聚类算法对线段进行聚类,每个类代表一种运动模式;最后,通过提出的二级检测算法判断待测轨迹是否异常。(2) 提出了一种基于LHOG和LHOF特征的人群异常检测算法。为了学习正常行为模式,首先结合灰度图像与光流图像对视频帧提取运动区域,然后在运动区域提取有效运动块,对运动块提取LHOG(Local Histogram of Oriented Gradient)和LHOF(Local Histogram of Optical Flow)特征,最后使用一类分类器建模得到正常行为模式的紧凑边界。检测过程中加入运动连续性约束,以抑制虚警噪声。(3) 提出了一种基于SVM的暴力异常检测算法。该算法首先根据光流场提取运动区域,然后对运动区域提取LHOG和LHOF两种底层特征,采用词包法把每个短视频的LHOG和LHOF特征投影为一个固定长度的向量,最后使用SVM分类器对特征向量进行分类,得到分类模型,用以检测短视频是否发生暴力行为。(4) 提出了一种基于深度学习的暴力异常检测算法。该算法采用卷积神经网络,基于行为识别框架构建了一个暴力检测网络—FightNet;以RGB图像、光流场图像和加速度图像三种模态数据为输入,提高了暴力检测的正确率;搜集了一个大型的暴力异常数据集,用以训练深度网络。(5) 提出了一种基于信息熵的人数估计算法。该算法首先提取运动区域,然后对运动区域进行透视校正,再对透视校正后的图像提取信息熵特征,得到信息熵与人数之间的回归模型,最后使用该模型预测人数。
Other AbstractThe abnormal behaviors are firstly characterized in this dissertation, and are divided into two categories: non-specific abnormal behaviors and specific abnormal behaviors. The crowd anomaly is studied for non-specific abnormal behaviors and the violence anomaly and the pedestrian number estimation are studied for specific abnormal behaviors. Moreover, a overview is made for different anomaly detection algorithms. Aiming at the different anomaly detection tasks and challenges, this dissertation shows different solutions. The main contributions of this dissertation are as follows: (1) A crowd anomaly detection algorithm based on the motion trajectory features is proposed. The research object is the tracking points obtained by different tracking methods. First, the trajectories are represented as different line segments according to the speed and angle of the trajectory points. Second, line segments are clustered into different classes using the DBSCAN clustering algorithm, and each class represents a motion pattern. At last, a two-level detection method is proposed to determine whether the trajectory is abnormal. (2) A crowd anomaly detection algorithm based on the LHOG and LHOF features is proposed. To learn the normal behavior patterns, the motion regions are firstly segmented according to the gray images and the optical flow images. Then the effective motion blocks are obtained in the motion regions, and the LHOG(Local Histogram of Oriented Gradient) and LHOF(Local Histogram of Optical Flow) features are extracted from the motion blocks. Finally, the one-class classifier is used to model the compact boundary of the normal behavior patterns. The motion continuity constraint is developed in the anomaly detection process to suppress false alarm noises. (3) A SVM-based violence anomaly detection algorithm is proposed. First, the algorithm extracts the motion regions according to the distribution of the optical flow field. Second, two kinds of low-level features, LHOG and LHOF are extracted from the motion regions. Third, the Bag of Words method is adopted to project the low-level features of each short video into a fixed length vector. Finally, the SVM classifier is used to classify the feature vectors. With the classification model, the short video is detected whether it is violent or not. (4) A violence anomaly detection algorithm based on deep learning is proposed. Using the convolutional neural network, a violence detection network, FightNet is constructed based on the behavior recognition framework. This network takes three kinds of modalities as the input data, i.e. RGB image, optical flow image and acceleration image, which increases the violence detection accuracy rate. Besides, a large set of violence anomaly dataset (VAD) is collected to train the deep network. (5) A pedestrian number estimation algorithm based on information entropy is proposed. First, the motion regions are extracted. Second, the motion regions are processed by perspective correction. Third, the information entropy features are extracted from the corrected images, and the regression model is constructed between the information entropy and the pedestrian number. Finally, the model is used to predict the pedestrian number.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/21830
Collection光电信息技术研究室
Affiliation1.中国科学院沈阳自动化研究所
2.中国科学院大学
Recommended Citation
GB/T 7714
周培培. 面向智能监控的人类异常行为检测方法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2018.
Files in This Item:
File Name/Size DocType Version Access License
面向智能监控的人类异常行为检测方法研究.(4872KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[周培培]'s Articles
Baidu academic
Similar articles in Baidu academic
[周培培]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[周培培]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.