Research On Gesture Recognition Based On ELM And WiFi Signals

With the development of human-computer interaction technology,the process from people adapting to computers to computers constantly adapting to people has been changed.Nowadays,human-centered interaction is the research focus of human-computer interaction technology.As one of the core technologies of human-computer interaction,gesture recognition has been widely used in sign language translation,smart home and virtual reality.At present,common gesture recognition technologies include wearable sensors and computer vision.Although these two methods can achieve high accuracy,they all have their own limitations.Gesture recognition based on wearable sensor is an active recognition method,which requires users to wear corresponding equipment,lacking convenience and high equipment cost.Vision-based gesture recognition is a passive recognition method,which does not require users to wear extra equipment,but its recognition accuracy is easily affected by external environment such as illumination and occlusion,which cannot guarantee personal privacy.Therefore,aiming at the problems of large energy consumption and difficult deployment of traditional gesture recognition methods,the gesture recognition method based on wireless sensor has high practical value.With the development of wireless communication technology and the wide deployment of WiFi devices,gesture recognition based on WiFi has become a research hotspot.In this paper,two hybrid deep learning models are proposed,which use channel state information(CSI)from WiFi devices for gesture recognition.The main research contents include: firstly,the original data are preprocessed by Butterworth filter,principal component analysis,wavelet transform and time-frequency analysis,and three different input parameters,namely the denoised CSI data,Doppler frequency shift component and CSI amplitude variance,are obtained as the input of the deep learning model.Secondly,on Widar data set,the de-noised CSI data and Doppler frequency shift component are used as input.The proposed CGRUELM deep hybrid model mines the spatio-temporal characteristics of two input parameters,thus realizing the recognition of six common human-computer interaction gestures.Finally,on the Sign Fi data set,taking CSI amplitude variance and Doppler shift component as input parameters,TCN-ELM model is proposed to extract the sequence features of the two input parameters,and realize the recognition of different sign language actions.From different input parameters,experimental environment,methods and data sets,the two deep learning models proposed in this paper are compared.The experimental results show that the two deep learning models proposed in this paper have achieved good accuracy on two different data sets and are effective classification models.