Research On Hand Gesture Recognition Method Based On CSI

With the development of human-computer interaction technology,gesture recognition has been widely used in daily life as the basic technology of intelligent home,sign language translation,virtual reality and other important Internet of things applications.At present,people generally use two gesture recognition technologies based on vision and wearable devices.However,hand gesture recognition technology based on vision is vulnerable to the influence of light conditions,high cost of equipment,unable to protect personal privacy and other issues.The other technology based on wearable equipment requires users to wear special equipment,which not only limits the freedom of users,but also is difficult to achieve human-computer interaction centered on people.Therefore,it is of great practical value and application significance to research a gesture recognition technology which does not need any equipment,is not affected by the light intensity,and has low deployment cost.Using channel state information from Wi-Fi devices for gesture recognition has many advantages like no need to carry devices,night work and low cost.It provides a new solution and technical method to eliminate the limitations of traditional gesture recognition technology.However,the high dynamic nature of wireless signal propagation,weak stability in the process of propagation and significant multipath effect in indoor environment are easy to limit the accuracy of gesture recognition.So this thesis proposes a hand gesture recognition method based on CSI.PEM was used to represent the variation of subcarrier fluctuation when the number of people in the room changes in order to reduce the limitation of gesture recognition in the experimental environment.Butterworth band-pass filter and principal component analysis were used to eliminate the high-frequency noise and impulse noise in the data,reduced the dimension of the data and the computational complexity in the experiment.Wavelet transform was used to extract the time-frequency features of different gestures.CNN-GRU model was designed to mine the spatiotemporal features of CSI sequence in order to realize dynamic gesture recognition.The experimental results show that the average recognition accuracy of HandFi is greatly improved in two different experimental environments.The recognition accuracy is 95.29%in the quiet office environment,while the recognition accuracy is 93.19%in the noisy laboratory environment.But the complexity of CNN-GRU model algorithm designed in this thesis was lower and the recognition accuracy could be achieved by using less data for training.It was found that the number of subcarriers,the size of gesture area and the transmission rate had a great impact on the recognition accuracy by the comparative experiments,which provided a basis for the later experimental environment settings.