WiTransformer: A Novel Robust Gesture Recognition Sensing Model with WiFi.

Area of Science:

• Computer Science
• Signal Processing
• Artificial Intelligence

Background:

• Human activity recognition (HAR) using WiFi signals offers non-invasive and ubiquitous sensing capabilities.
• Existing HAR systems often struggle with increased task complexity, such as numerous action classes or signal distortions.
• Previous research primarily focused on model precision, neglecting the impact of task complexity on performance.

Purpose of the Study:

• To develop a robust human activity recognition system using WiFi signals that can handle complex tasks.
• To introduce a novel approach based on pure Transformers, overcoming limitations of conventional convolutional and recurrent models.
• To investigate the effectiveness of modified Transformer architectures for WiFi-based gesture recognition.

Main Methods:

• Proposed WiTransformer, a novel architecture based on pure Transformers, eliminating conventional backbones.
• Introduced the Body-coordinate Velocity Profile, a cross-domain WiFi signal feature, to lower Transformer pretraining thresholds.
• Developed two modified Transformer architectures: united spatiotemporal Transformer (UST) and separated spatiotemporal Transformer (SST).

Main Results:

• UST achieved 86.16% accuracy on the most complex dataset (TDSs-22), outperforming other backbones.
• UST demonstrated remarkable robustness, with accuracy decreasing by at most 3.18% as task complexity increased.
• SST models failed due to insufficient inductive bias and limited training data, as predicted.

Conclusions:

• WiTransformer, particularly the UST architecture, offers a robust solution for WiFi-based human activity and gesture recognition.
• The Body-coordinate Velocity Profile effectively enables Transformer models to perform well on WiFi HAR tasks with smaller datasets.
• The study highlights the potential of Transformer architectures for HAR while emphasizing the importance of model design for task complexity.