Fine-grained Action Recognition Based On Wi-Fi Channel State Information

With the continuous development of wireless communication technology and the rapid popularization of Wi-Fi devices,Wi-Fi signals exist in people's daily work and living places.Wi-Fi signal can not only transmit communication data,but also identify human activity by extracting channel state information(CSI)data from Wi-Fi signal.The principle is that human activity can affect the CSI of Wi-Fi signal.The activity can be recognized by studying the changing pattern of CSI.Compared with traditional image-based or sensor-based recognition methods,the CSI-based methods use Wi-Fi signals that are already widely available in the environment for recognition.It has the advantages of non-line-of-sight perception,non-contact perception,low cost,wide working range,no invasion of privacy.In recent years,there have been a great number of applications using CSI of WiFi for recognition,including daily behavior recognition,gait recognition,gesture recognition,population statistics,fall detection,etc.Because Wi-Fi channel state information is fine-grained physical information,which is sensitive to environmental changes,it has great advantages in fine-grained activity recognition.Therefore,this thesis proposed two gesture recognition methods based on CSI,which can recognize six fine-grained gestures: upward,downward,up and down,left,right and left swings.The main work content of this thesis are as follows:1.CSI gesture recognition method based on CRNNAiming at the problem that the traditional work relies on manually designed features to select subcarriers,a new subcarrier selection method is proposed.Through the result of self-organizing map(SOM)clustering to select the best subcarrier of CSI.In the feature extraction stage,instead of using statistical features,spectral feature and other common features,we use the Gramian Angular Summation Fields(GASF)method transforms one-dimensional CSI time series into two dimensional images as training data.In this way,we can not only use the deep neural network to extract features,but also encode data sequences with different lengths into images of the same size,reducing the steps of sequence segmentation.Finally,training Convolutional Recurrent Neural Network(CRNN)for gesture classification.The experimental results show that this method can not only recognize gestures without manually constructing Wi-Fi propagation model,design sub-carrier selection features and classification features,but also achieves high recognition accuracy.2.An end-to-end CSI gesture recognition methodMost of the current methods divide the whole recognition process into data preprocessing,feature construction,training and classification stages,and professional domain knowledge is required to design the methods used in each stage.In addition,the error of the former stage may affect the later stage,resulting in error accumulation,which makes the whole system unable to achieve the best performance.To solve the above problems,an end-to-end identification method is proposed.After low-pass filtering,the original CSI data can be input into the model,and the data on all subcarriers can be used for training and prediction.Two branches are designed in the end-to-end model.One is Bi-directional Long Short-Term Memory(Bi-LSTM)with Attention mechanism,is responsible for extracting the time information in the CSI sequence change pattern.The other is the Multi-scale Convolutional Neural Network(MCNN)which adds the Squeeze-and-Excitation Block(SE Block)to extract spatial and frequency features from the multi-scale input data of CSI.In addition,the Connectionist Temporal Classification(CTC)loss function used in the training process,so that the original CSI data sequence does not need to intercepted into equal length fragments.The experimental results show that the proposed method not only has a significant improvement in accuracy than traditional methods that rely on manual designed features,but also reduces the complexity of the recognition system.