Device-Free Human Behaviour Recognition is automatically recognizing physical
activity from a series of observations, without directly attaching sensors to the subject.
Behaviour Recognition has applications in security, healthcare and smart homes. The
ubiquity of WiFi devices has generated recent interest in Channel State Information
(CSI) that describes the propagation of RF signals for behaviour recognition, leveraging
the relationship between body movement and variations in CSI streams. Existing work
on CSI based behaviour recognition has established the efficacy of deep neural network
classifiers, yielding performance that surpasses traditional techniques. In this paper,
we propose a deep Recurrent Neural Network (RNN) model for CSI based Behaviour
Recognition that utilizes a Convolutional Neural Network (CNN) feature extractor with
stacked Gated Recurrent Units (GRUs) for sequence classification. We also examine CSI
de-noising techniques that allow faster training and model convergence. Our model has
yielded significant improvement in performance compared to existing techniques. In
addition to this, we propose a novel one-shot learning approach for behaviour recognition
using CSI data, with very limited training samples.