Utilizing deep learning models in CSI-based human activity recognition.

Area of Science:

• Computer Science
• Electrical Engineering
• Biomedical Engineering

Background:

• Channel State Information (CSI) from WiFi 802.11n is a novel data source for human activity recognition (HAR).
• HAR has critical applications in areas like remote patient monitoring.
• Existing HAR methods often struggle with accuracy and feature extraction.

Purpose of the Study:

• To develop and evaluate deep learning models for enhanced human activity recognition using WiFi CSI data.
• To compare the performance of four distinct deep learning architectures for HAR.
• To demonstrate the effectiveness of CSI-based HAR for various daily activities.

Main Methods:

• Four deep learning models were designed: CNN-GRU, CNN-GRU with attention, CNN-GRU-CNN, and CNN-LSTM-CNN.
• Models were trained using CSI amplitude data collected via a CSI tool.
• Model performance was evaluated using a 70% training and 30% testing data split, including unseen data.

Main Results:

• The CNN-GRU and CNN-GRU with attention models achieved the highest average recognition accuracies at 99.31% and 99.16%, respectively.
• Models demonstrated robust performance on unseen data, reaching nearly 100% accuracy for most activities.
• The CNN-GRU model achieved 99.46% accuracy for lying down detection, showcasing high precision.

Conclusions:

• Deep learning models effectively leverage WiFi CSI data for accurate human activity recognition.
• The proposed models, particularly CNN-GRU and CNN-GRU with attention, offer superior performance compared to existing HAR approaches.
• CSI-based HAR systems show promise for applications requiring reliable human movement monitoring.