Research On Gesture Recognition Method Based On Channel State Information

Recently,gesture recognition technology has been widely used in smart home,intelligent driving,intelligent wearing,smart healthcare and so on.However,sensorbased recognition method requires users to wear the equipment strictly according to the requirements,which brings inconvenience to them in daily life.Computer visionbased gesture recognition has high dependence on lighting conditions and can easily threaten the users’ privacy.Therefore,Wi Fi-based gesture recognition technology is a better choice than the others due to its advantages such as low deployment cost,users’ privacy protection and inflexibility to light intensity,which makes Wi Fi-based gesture recognition technology show great potential development and commercial value.This thesis is based on Channel State Information(CSI)of Wi Fi and combines the efficient physical model and deep neural network to realize the gesture recognition of indoor personnel.This thesis solves the problems of current sensing model such as inaccurate gesture estimation,low feature-extracting efficiency and difficulty of extending gesture recognition technology to mobile.The main work is as follows:(1)To accurately estimate gesture changes,a gesture recognition model based on CSI is derived in this thesis.Firstly,a CSI Fresnel zone model focusing on the transceiver is established according to the propagation characteristics of Wi-Fi signal.Then,a CSI quotient model is used for eliminating the phase error and most ambient noise,so the dynamic changes of gestures are transformed into the changes of CSI quotient in the complex plane.Finally,the Hampel filter is adopted to eliminate the abnormal values.It also eliminates the interference of the hardware noise with the combination of Discrete Wavelet Transform(DWT).The results show that the model reveals the exact relationship between the subtle motion of the finger and the form of CSI fluctuation.The experimental results show that the gesture recognition model reveals the exact relations between the subtle motion of the finger and the form of CSI fluctuation.(2)To achieve precise gesture recognition,this thesis extracts the feature values which can accurately describe the gesture class from the subcarrier information of all communication links by using the established deep neural network.Firstly,the gesture samples are normalized and the input formats are unified.Then,the pre-processed CSI information is input into the multiscale convolutional network to extract the spatial features.At the same time,a cross-channel interactive attention mechanism is used to optimize the feature extraction process.Finally,LSTM is used to collect the time-domain features while the codec structure is used to filter the features.The feature extraction ability of the model is improved by combining the self-attention mechanism and short-circuit connection.The experimental results show that this method can achieve 96.7% recognition accuracy.(3)To extend gesture recognition technology to mobile devices,this thesis proposes a gesture recognition method based on knowledge distillation.First,principal component analysis(PCA)and local outlier factor algorithm are used to adaptively segment dynamic gestures to eliminate redundant static components.Then,the spatiotemporal features of continuous gesture images are mined by using the Efficient Net-GRU model.Finally,knowledge transfer is accomplished by model distillation to improve the recognition accuracy of student model.Experiments show that this method can make the recognition accuracy of the lightweight student model reach 94.2%,and effectively improve the efficiency of gesture detection.This thesis contains 70 figures,11 tables and 85 references.