Multi-Feature Unsupervised Domain Adaptation (M-FUDA) Applied to Cross Unaligned Domain-Specific Distributions in...

Area of Science:

• Computer Science
• Electrical Engineering
• Signal Processing

Background:

• Human-computer interaction (HCI) relies on human activity recognition (HAR) for innovation.
• Traditional HAR methods raise privacy concerns due to infrastructure and user cooperation needs.
• Wi-Fi sensing using channel state information (CSI) offers privacy-preserving, device-free HAR but faces challenges with new users, environments, and scalability.

Purpose of the Study:

• To develop an unsupervised multi-source domain adaptation technique for robust HAR.
• To address limitations of existing Wi-Fi-based HAR systems in cross-domain scenarios.
• To enhance the adaptability and scalability of device-free sensing.

Main Methods:

• Proposed an unsupervised multi-source domain adaptation technique.
• Aligned diverse data distributions with target domain variations (new users, environments, atmospheric conditions).
• Utilized a preprocessing module to convert CSI data into image-like formats for analysis.

Main Results:

• Achieved significant improvements over baseline methods in cross-domain HAR tasks.
• Demonstrated an average micro-F1 score of 81% for cross-user tasks.
• Showcased 76% for cross-user and cross-environment, and 73% for cross-atmospheric tasks.

Conclusions:

• The proposed method effectively enhances system adaptability for Wi-Fi-based HAR.
• The approach is suitable for scalable, device-free sensing in realistic cross-domain scenarios.
• This technique overcomes limitations related to unseen users, new environments, and varying atmospheric conditions.