| With the development of China’s social economy,the phenomenon of social aging is becoming more and more intense.In non disease death,fall accidents have become an important cause of death for the elderly.Therefore,the health security and nursing for the elderly living alone need to be strengthened.Therefore,it is of great significance to detect the occurrence of fall events in time.Aiming at the problems that the existing wearable device based schemes are uncomfortable to wear,the camera scheme is easy to be blocked,and benefiting from the popularity of WiFi devices,this paper implements a fall detection scheme based on WiFi devices.In this scheme,the human posture is predicted first,and then the fall is judged.In order to mark the WiFi signal in real time,this scheme uses the camera as an auxiliary way to obtain the video image data synchronized with the WiFi signal,and then uses Alphapose to obtain the human skeleton information from the image.Finally,this scheme can use the neural network Dresnet to establish the mapping relationship between the WiFi signal and the human skeleton information.Because the change of human posture causes the change of WiFi signal,the information corresponding to the change of human key point coordinates needs to be retained when extracting the WiFi signal,and there is a relationship between the key point coordinates,which requires that the neural network used in this scheme needs to have a large range of receptive field and retain the change information.Therefore,this scheme uses dilated convolution to realize the above functions.The main contents and contributions of this paper are as follows:1)Completed the cross modal acquisition and processing mechanism combining of WiFi signal and image information.In order to mark the WiFi signal more real-time and intuitively,a cross modal signal system is built using WiFi equipment and camera.Both work at the same time to collect the wireless WiFi signal and video signal corresponding to the human posture.Cross modal provide more information,such as space,time,frequency,attitude information and so on.Cross modal information helps to improve the robustness and generalization of fall detection system.2)The Dilated residual neural network(Dresnet)used in this scheme is composed of Dilated Convolution and Residual Neural Network(Resnet).The dilated convolution can have a large sensing range without losing resolution.The pyramid structure can avoid the problem that the dilated convolution only considers the long-distance relationship and ignores the short-distance relationship.Therefore,the pyramid dilated convolution is selected as the feature extraction network in this scheme;Moreover,once the neural network deepens,the gradient may disappear or the gradient may explode.In order to avoid this situation,the residual neural network Resnet is selected as the training network of characteristic data in this scheme;Finally,the Dresnet neural network used in the cost scheme is formed.3)Implemented a fall detection system.After the original signal is collected,preprocess the collected original signal,and then use the human body posture information obtained from the video signal to supervise the training of the Dresnet neural network model on the WiFi signal.The trained model can only detect the wireless WiFi signal.Predict to obtain human body posture information,and then use a two-class SVM classifier to judge whether it is in a fall state according to the human body posture information predicted by the model.Experiments show that the detection accuracy of the key points of human posture obtained by this scheme can reach 88.3%,and the accuracy of fall detection can reach 92.9%. |
