Location Independent Gesture Recognition Using Channel State Information

Abstract

Gesture recognition has been the subject of intensive research in recent years owing to its wide applications. Unlike traditional systems, which usually require wearable sensors, many recent works have achieved the desirable gesture recognition performance using wireless channel state information from commercially available WiFi devices. However, existing works generally require training new models for different locations due to the location-dependent nature of channel state information. This paper proposes a location-independent system that can recognize gestures performed in a new location without training a new model. Our approach uses disentanglement that extricates location and other extraneous information from those needed for gesture recognition. The implementation is based on an unsupervised invariance induction framework consisting of feature extraction, a multi-output latent space, gesture recognition, and decoder modules. The key idea in designing this system is to separate gesture-dependent features from location-dependent features. Specifically, the feature extraction module consisting of a long short-term memory network is employed to select representative features; it essentially serves as an encoder to generate the latent space. During the training process, the network learns to cluster features representation for the gesture recognition and decoder by minimizing the total loss of the gesture recognition and decoder modules. We test our system with a dataset collected from various subjects performing four different gestures in multiple locations in seven rooms with different layouts. The results show that our location-independent gesture recognition system can achieve 88.69% accuracy for new locations.