基于深度学习和WebRTC的智能跌倒监控系统研究

发布时间：2018-11-28 10:58

【摘要】：随着我国人口老龄化程度不断加剧和社会生活压力的日渐增长,空巢老人的比例不断上升。老人独自居住造成很多社会问题,尤其是老人的健康问题,其中老人意外跌倒损伤是影响老人健康的主要原因之一。老人跌倒后没有被及时处理会增加二次伤害,甚至导致老人意外死亡。据相关统计50%以上的老人跌倒发生在家中,如果能够对老人的跌倒行为进行实时的监控,当老人发生跌倒行为后能准确识别,并向其监护人发送提醒信息,使老人得到及时的救助将极大减少跌倒对老人健康的伤害。基于此,本文展开了基于深度学习和WebRTC的智能跌倒监控系统的研究。主要工作如下:在分析基于深度学习和WebRTC的智能跌倒监控系统的功能需求的基础上,提出了系统的技术方案。该方案使用深度学习技术实现老人跌倒行为的智能识别,基于WebRTC视频传输架构实现远程视频传输。展开了基于深度学习的跌倒行为识别的研究。首先,提出并仿真实现了基于视频帧和VGGNet-16卷积神经网络模型的跌倒识别方法;对Le2i、SDU、UCF-101开源视频数据集,进行水平翻转、对比度和亮度调节、加噪等数据增强处理后构建卷积神经网络的训练集和测试集,对上述方法进行训练和测试。实验结果表明:该方法识别结果强烈依赖于训练场景。其次,针对上述的问题,提出了一种基于双流卷积神经网络的跌倒识别方法。该方法:一路采用场景相减法检测视频中的运动目标,将视频帧中运动目标加框标记后输入到3D-CNN模型中进行跌倒识别;另一路采用光流法提取视频的光流图,将光流图输入到VGGNet-16模型中进行跌倒识别;最后将两路模型的跌倒识别结果进行线性加权融合。实验结果表明:基于双流卷积神经网络的跌倒识别方法跌倒识别率为96%,比基于视频帧和VGGNet-16卷积神经网络模型的跌倒识别方法识别率提高了51%,比基于运动目标检测和3D-CNN的跌倒识别方法提高了4%,比基于光流图和VGGNet-16的跌倒识别方法提高了3%,且有良好的泛化能力。展开了基于WebRTC的远程视频监控的研究。提出了基于WebRTC的远程视频监控方案,搭建了视频传输的信令服务器和穿网服务器,并基于WebRTC实现了视频采集端、远程视频监控端。通过搭建含NAT的网络实验环境对系统进行测试,实验结果表明:本文实现的信令服务器、穿网服务器工作正常,视频传输客户端功能正常,且能够穿越防火墙和NAT的限制实现P2P的视频传输。
[Abstract]:With the aggravation of the aging population and the increasing pressure of social life, the proportion of empty nest elderly is increasing. Living alone causes many social problems, especially the health problems of the elderly, among which the accidental fall and injury of the elderly is one of the main reasons affecting the health of the elderly. Failure to be dealt with in time after a fall can increase secondary injuries and even lead to accidental death. According to related statistics, more than 50% of the elderly fall at home. If the fall behavior of the elderly can be monitored in real time, it can be accurately identified when the fall behavior occurs, and send a warning message to their guardian. Timely assistance to the elderly will greatly reduce the harm of falls to the health of the elderly. Based on this, the research of intelligent fall monitoring system based on deep learning and WebRTC is carried out in this paper. The main work is as follows: based on the analysis of the functional requirements of the intelligent fall monitoring system based on deep learning and WebRTC, the technical scheme of the system is proposed. This scheme uses depth learning technology to realize the intelligent recognition of fall behavior of the elderly, and realizes remote video transmission based on WebRTC video transmission architecture. The research on recognition of fall behavior based on deep learning is carried out. Firstly, a fall recognition method based on video frame and VGGNet-16 convolution neural network model is proposed and simulated. The training set and test set of convolutional neural network are constructed after the Le2i,SDU,UCF-101 open source video data set is flipped horizontally, the contrast and brightness are adjusted, and the noise is increased. The above methods are trained and tested. The experimental results show that the recognition results are strongly dependent on the training scene. Secondly, a fall recognition method based on two-flow convolution neural network is proposed. This method uses scene subtraction method to detect moving target in video frame, then marks moving object in video frame, then inputs it into 3D-CNN model for fall recognition. On the other hand, the optical flow image of video is extracted by optical flow method, and the optical flow graph is input into the VGGNet-16 model for fall recognition. Finally, the fall recognition results of the two models are fused linearly. The experimental results show that the fall recognition rate based on two-stream convolution neural network is 96, which is 51% higher than that based on video frame and VGGNet-16 convolution neural network model. Compared with the moving target detection and the fall recognition method based on 3D-CNN, the proposed method has a better generalization ability than the optical flow graph and the fall recognition method based on VGGNet-16. The research of remote video surveillance based on WebRTC is carried out. A remote video surveillance scheme based on WebRTC is proposed. The signalling server and the netting server of video transmission are built. The video collection terminal and remote video monitoring terminal are implemented based on WebRTC. The experimental results show that the signaling server, the network piercing server, and the video transmission client function are working well, and the video transmission client function is normal. And can pass through the firewall and NAT restrictions to achieve P2P video transmission.
【学位授予单位】：华东交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP18

【参考文献】