Gesture Recognition Based On Channel State Information Of WiFi

With the population of application of WiFi devices,Wi-Fi not only provides the convenient and high-speed data service,but shows the great potential on the object detection and recognition.In the past several years,the research on human motion detection based on WiFi has drawn much attention to many scholars due to the fewer restrictions of atmosphere of detection and no necessity of wearing expensive devices.Gesture recognition,one of branch of human motion detection,require more precise detecting method to achieve it because of smaller scale of motion.So far,the preferable method of gesture recognition is collecting the channel station information of WiFi from the commercial network card although this method shows low accuracy and is easy to bring extra noise leading to the problem of data processing.On the other hand,the algorithm of recognition is simple and hard to achieve the high accuracy in the NLOS.Referring to the problems above,this thesis provide the better solution.The USRP device is applied as a receiver for its high-precision and the WiFi receiver algorithm is designed and modified on the platform of GNU Radio to collect CSI data which is added timestamp.After sample alignment,the PCA algorithm is applied to reduce the white noise and discrete wavelet transform is used to extract the features.In order to obtain the optimal algorithm,Five machine learning algorithms are applied to classify a set of eight gestures under the ideal environment.The result shows the optimal algorithm is the random forest.Moreover,the feature selection is utilized to promote the accuracy of gesture recognition.Finally,for the sake of validating the robustness and find the scope of application of the thesis,the different environment is set to carry the experiment such as difference of CSI sampling rate,number of interferences,distance of receiver and emitter,and NLOS scenario.The result shows that the random forest algorithm can achieve the average accuracy of 96% in the LOS scenario and 92% in the NLOS.More importantly,it can reach the average accuracy of more than 90% in the atmosphere of low sampling rate of CSI at 256 kpts/s,which throw the light on the low demanding of WiFi transmission.within five members of interference and distance at less than 4.5 m between the receiver and emitter can also achieve the average accuracy of more than 90%.The result proves the robustness of the gesture recognition based on CSI of WiFi this thesis proposed as well as the scope of application.