经过深入研究近年来发展迅速的深度学习技术, 并学习卷积神经网络处理视频数据的方法本文, 在传统3D卷积神经网络的基础上改进了网络结构.同时, 考虑到考场采用双摄像头监控系统, 可从不同视角观察考生的考试行为, 本文提出了基于双路的考场异常行为识别方法.该方法结合了改进的3D卷积神经网络和双摄像头的监控系统, 设计了新的双路网络结构的视频特征提取器, 可以提取不同视角下的考生行为特征, 并将双路网络提取到的行为特征向量进行融合.通过提取正常考试行为的特征向量, 在LibSVM中训练出考场行为的分类器, 该分类器可以对测试视频的特征向量进行分类, 由此判断测试视频中是否存在异常行为.该方法使用双路视频特征进行异常识别, 在考场行为数据集中有着较高的识别正确率.%After a profound research on deep learning technology developed rapidly in recent years and a comprehensive study on the methods of processing the video data with the convolution neural networks, a network structure is improved on the basis of the traditional 3D convolution neural network in this paper. Moreover, considering the fact that the student behavior could be observed from two different perspectives because of the monitoring system in the examination room containing two independent cameras, an abnormal behavior recognition method based on dual stream videos is proposed, where the improved 3D convolution neural network and the dual-camera monitoring system are analyzed simultaneously for designing a novel video feature extractor to extract the behavior characteristics of the students in two perspectives separately. Then, a classifier of behavior in examination with the capacity of sorting the behavior characteristics in the test video and detecting the abnormal behaviors is trained in the LibSVM by extracting the characteristics of normal behavior in examinations. This method utilizes the characteristics of dual steams video to realize the abnormal behavior recognition, exhibiting high recognition accuracy in the behavior data set.
展开▼
