Direction-agnostic gesture recognition system using commercial WiFi devices

Abstract

In recent years, channel state information (CSI) has been used to recognize hand gestures for contactless human–computer interaction. However, most existing solutions require precision hardware or prior learning at the same angle both during training and for inference/training in order to achieve high recognition accuracy. This requirement is unrealistic for practical instrumentation, where the orientation of a subject relative to the RF receiver may be arbitrary. We present direction-agnostic hand gesture recognition utilizing commercial WiFi devices to overcome low accuracy in non-trained observation angles. To achieve equal conditions in all recognition angles, first of all, through the circular antenna arrangement to mitigate the impact of user direction changes. Then, the orientation of users is estimated by the Fresnel zone model. Finally, the feature mapping model of users in different orientations is established, and the gesture features in the estimated direction are mapped to the benchmark direction to eliminate the influence caused by the change of user orientation. Experimental results in a typical indoor environment show that WiNDR has superior performance, and the average recognition accuracy of five common gestures is 92.38%.