| With the rapid development of Internet of Things and Artificial Intelligence technology,human behavior sensing has been applied more and more in security monitoring,smart home,human-computer interaction and other fields,and its related technologies have become one of the research hotspots.Compared with traditional technologies,such as computer vision and special sensors,behavior sensing technology based on Wi-Fi signal has the advantages of device-free,lower cost and better privacy protection,which has attracted wide attention of researchers in recent years.However,in the actual scene,due to the complexity of the environment and the high dynamics of wireless signals,the detection accuracy of behavior sensing model based on Wi-Fi signals often decreases.Therefore,the key to realize Wi-Fi behavior sensing is how to effectively obtain the data reflecting human behavior and design a recognition model with high recognition accuracy and strong generalization ability.The main research contents of this thesis are as follows:(1)Aiming at the problems of insufficient feature extraction and low recognition accuracy of existing recognition models based on Wi-Fi signal behavior sensing,in this thesis,we proposed a novel CSI data expression method---CSI spectrum map.Different from time series data,CSI spectrum map can show the frequency changes of different behaviors of human body,and can make full use of the advantages of convolution neural network in image feature extraction to improve the accuracy of behavior recognition.Experimental results show that the CSI spectrum map can make the recognition accuracy of the model as high as 99.71%.(2)Aiming at the problems of training difficulty and low real-time performance caused by the large number of parameters in the current deep neural network model for human behavior recognition,in this thesis,we proposes a new deep neural network model---Greed Net.The model takes CSI spectrum map as input,uses asymmetric convolution kernel,and uses multi-scale feature extraction technology to capture more levels of image features,and then effectively fuses the features to improve the accuracy of human behavior recognition.Experimental results show that Greed Net not only has fewer parameters,but also has higher accuracy and robustness compared with existing human behavior recognition models.In summary,in this thesis,we proposed a novel CSI data expression method---CSI spectrum map,which realizes the accurate extraction of human behavior characteristics.Greed Net,a multi-scale and lightweight neural network model for human behavior recognition with spectrum map as input,is constructed to ensure that the model parameters can be reduced as much as possible with high accuracy. |
