Abstract: Aiming at the problems of high energy consumption and difficult deployment in traditional gesture recognition methods , a WiFi-based gesture recognition method is proposed. By extracting the Doppler frequency shift components from the fine-grained channel state information collected from WiFi signal, the problem of unclear mapping relationship between the statistical features extracted by wireless gesture recognition method and specific gestures is solved. Meanwhile, a deep hybrid model of CGRu-ELM is proposed to extract and classify the extracted Doppler frequency shift components, and to recognize six commonly used human-computer interaction gestures. The experimental results show that the average 
accuracy of this method for gesture recognition with WiFi signal as input parameter is 93.4%.

Key words: channel state information, Doppler frequency shift, gesture recognition, deep learning