Device-Free Multi-Location Human Activity Recognition Using Deep Complex Network.

Area of Science:

• Computer Science
• Signal Processing
• Artificial Intelligence

Background:

• Wi-Fi-based human activity recognition offers device-free, privacy-preserving sensing, unaffected by lighting conditions.
• Current AI-driven methods show high accuracy but struggle with multi-location recognition due to signal variations.
• Existing solutions require extensive data collection at diverse locations, proving labor-intensive.

Purpose of the Study:

• To develop a novel method for accurate multi-location human activity recognition using Wi-Fi signals.
• To address the challenge of location variations impacting wireless sensing performance.
• To reduce the need for extensive data samples at each location.

Main Methods:

• An amplitude- and phase-enhanced deep complex network (AP-DCN) was developed to utilize both signal components for richer information extraction.
• A deep complex network-transfer learning (DCN-TL) approach was proposed to handle unbalanced sample numbers and enable knowledge sharing across locations.
• Experiments were conducted in an office environment with 24 distinct locations and five different human activities.

Main Results:

• The AP-DCN method achieved a recognition accuracy of 96.85%.
• The DCN-TL method demonstrated a recognition accuracy of 94.02%.
• Both methods effectively utilized amplitude and phase information, improving recognition from limited data.

Conclusions:

• The proposed AP-DCN and DCN-TL methods offer a robust solution for multi-location human activity recognition.
• These techniques significantly enhance sensing accuracy by effectively leveraging wireless signal characteristics and transfer learning.
• The findings demonstrate a practical and efficient approach to device-free activity recognition in complex environments.