Attention-Based Gait Recognition and Walking Direction Estimation in Wi-Fi Networks

Most existing Wi-Fi-based gait recognition systems consider gait cycle detection as a critical process. However, the noise mixed in dynamic measurements obtained from commercial Wi-Fi devices makes it hard to detect gait cycles. Herein, we adopt the attention-based Recurrent Neural Network (RNN) encoder-decoder and propose a cycle-independent human gait recognition and walking direction estimation system, termed AGait, in Wi-Fi networks. For capturing more human walking dynamics, two receivers together with one transmitter are deployed in different spatial layouts. The Channel State Information (CSI) from different receivers are first assembled and refined to form an integrated walking profile. Then, the RNN encoder reads and encodes the walking profile into primary feature vectors. Given a specific gait or direction sensing task, a corresponding and particular attention vector is computed by the decoder and is finally used to predict the target. The attention scheme motivates AGait to learn to adaptively align with different critical clips of CSI data for different tasks. We implement AGait on commercial Wi-Fi devices in three different indoor environments, and the experimental results demonstrate that AGait can achieve average F-1 scores of 97.32 to 89.77 percent for gait recognition from a group of 4 to 10 subjects and 97.41 percent for direction estimation from 8 walking directions.