Design And Implementation Of Multiple Human Indoor Environmental Monitoring System Based On WiFi Channel State Information

The ubiquity of WiFi devices and the academic community for extensive researches on wireless sensing enable us to sense environmental changes in a Device-Free manner,which is significant to the development of Smart Home,IoT and other fields.However,most of the previous WiFi sensing applications focus on single human environment,which greatly hinder this technology from real implementation.The realization of how many targets or in other words crowd counting is the first step-stone for intelligent sensing in a multiple human environment and has profound significance.Therefore,we propose DeepCount-a system using Channel State Information(CSI)to infer the number of people in multiple human environment which contains two parts:1.Crowd counting model based on Deep Learning techniques.In this stage,DeepCount uses powerful learning ability of neural network to fit the relationship between CSI amplitude,phase information and the number of people and constructs the crowd countiong model with the help of deep learning approaches.It involves the analysis of the basic characteristics of CSI data,the use of Butterworth and Weighted Moving Average Filter to remove CSI amplitude noises,the research of CSI phase sanitization algorithm,the use of fully connected neural network to construct crowd counting model with deep learning approaches.Finally,the recognition accuracy of crowd counting model is 82.3%.2.Crowd counting model based on door switch with error correction mechanism.This stage attempts to explore how to further improve the accuracy of recognition.DeepCount proposes the mechanism with error correction to correct the result predicted by the crowd counting model with the help of human activity recognition model.The construction of activity recognition model including the use of Discrete Wavelet Transform(DWT)to extract features,the use of Hidden Markov Model(HMM)to construction the model,the error correction mechanism to retrain the weights of the last layer of neural network to correct the results.Experiments show that DeepCount can further improve the accuracy of counting model and achieves the accuracy of 87% at last.