Time Series Focused Neural Network for Accurate Wireless Human Gesture Recognition

Wireless human gesture recognition has spawned many emerging applications, e.g., intelligent control and sign language communications based on WiFi. By treating wireless signals as images, existing WiFi-based gesture recognition methods adopt 2-Dimensional Convolutional Neural Networks (2D CNNs) to capture the feature of transitions on channels. Although 2D CNNs are practical, such approaches need to pay more attention to the information in the temporal dimension, i.e., the channel of time-series features. To better exploit the information in the temporal dimension, we propose a time-series focused gesture recognition approach using WiFi signals named T-SignFi. Our proposed T-SignFi consists of a preprocessing AdaNorm normalization and a CNN-based 3D sign language recognition model to pay sufficient attention to the features in both temporal and channel dimensions. Experiment results based on the real-world dataset show that, compared to the four Sota methods, our method has significant advantages on all three datasets. Improvements of 4.84%, 3.31%, 3.42%, and 2.73% were achieved on the home and lab data, respectively. Our proposed T-SignFi could also effectively improve recognition performance with high generalization ability. In addition, for a challenging case, i.e., actions combining some body parts, including finger, arm, hand, and head movements, the proposed T-SignFi can also recognize them with high accuracy.