Human Activity Recognition Through Augmented WiFi CSI Signals by Lightweight Attention-GRU.

Area of Science:

• Computer Science
• Electrical Engineering
• Signal Processing

Background:

• Human Activity Recognition (HAR) is crucial for various applications.
• Existing state-of-the-art (SOTA) HAR models often have large computational and memory footprints.
• WiFi Channel State Information (CSI) offers a promising, non-intrusive sensing modality for HAR.

Purpose of the Study:

• To develop a lightweight and efficient model for HAR using WiFi CSI.
• To maintain high recognition accuracy while drastically reducing model size and computational cost.
• To address the limitations of current SOTA HAR models in terms of complexity.

Main Methods:

• Utilized a single-layer Gated Recurrent Unit (GRU) with an attention mechanism for HAR.
• Implemented data augmentation and pruning techniques to reduce model complexity.
• Evaluated the proposed model on four diverse datasets, including the ARIL dataset.

Main Results:

• Achieved high accuracy, reaching approximately 98.92% on the ARIL dataset.
• Significantly reduced model size from 252.10 million to 0.0578 million parameters.
• Drastically decreased computational cost from 18.06 GFLOPs to 0.01 GFLOPs.
• Outperformed a SOTA model on the ARIL dataset in terms of accuracy and efficiency.

Conclusions:

• The proposed lightweight GRU-based model with attention is highly effective for HAR using WiFi CSI.
• The model offers a significant reduction in complexity without compromising performance.
• This approach is well-suited for practical, resource-constrained HAR deployments.