摘要
为了解决当前视频监控系统对异常行为检测精度与自我学习能力较弱的问题,提出了基于光流场分析与深度学习的视频监控方法.首先,引入光流场检测算法,利用图像序列中目标像素的强度数据时域变化来确定运动行为是否异常,从而建立视频目标行为识别算子,获取异常行为光流特征;并利用卷积神经网络对光流特征进行逐层训练,设计自我学习机制,增强系统对异常行为的检出率;最后,基于.NET平台与Accord开源库,对本文监控系统进行实现.实验测试结果显示:与当前视频监控系统相比,本文算法拥有更高的异常行为检出力和深度学习升级能力.
In order to find a solution to the detecting accuracy of current video monitoring system which de-tects the abnormal behavior and the problem of self-learning ability, the method of video monitoring system based on optical flow field analysis and in-depth study is proposed. First of all, the detection algorithm opti-cal flow field is introduced, to make sure whether the movement behavior is abnormal which takes advantages of the time-domain change of intensify data of target pixel, then to built video behavior identification opera-tor thus the optical flow characteristics of abnormal behaviors are obtained; convolutional neural network is used to train the optical flow characteristics layer by layer and the self-learning mechanism is devised to en-hance the relevance ratio of abnormal behaviors, finally the video monitoring system based on the. net plat-form and Accord open source library is realized. The experimental results show that : compared with the cur-rent video system, the algorithm in this paper acquires higher ability of detecting the abnormal behaviors and deeply learning to upgrade.
作者
刘勇
Liu Yong(Chuzhou Branch, Anhui Open University,Chuzhou 239000,Chin)
出处
《湘南学院学报》
2017年第2期18-23,共6页
Journal of Xiangnan University
关键词
光流场
深度学习
视频监控
特征训练
卷积神经网络
optical flow field, deep learning, video monitoring, characteristics of training, convolutional neural network