Robust Indoor Human Activity Recognition Using Wireless Signals.

Area of Science:

• Computer Science
• Electrical Engineering
• Human-Computer Interaction

Background:

• Vision-based activity recognition faces limitations due to occlusions, viewpoint changes, and lighting variations.
• Wireless signals, specifically Wi-Fi, offer a complementary sensing modality for activity detection.
• Channel State Information (CSI) from Wi-Fi signals contains rich patterns of human movement.

Purpose of the Study:

• To develop a robust indoor human activity recognition framework using only one pair of Wi-Fi transmission points (TP) and access points (AP).
• To explore the properties of CSI for reliable activity detection.
• To create a system insensitive to location, orientation, and speed variations.

Main Methods:

• Selected indoor human actions as primitive actions for a training set.
• Employed an online filtering method to smooth CSI curves and retain pattern information.
• Utilized a segmentation method to isolate primitive action patterns from MIMO signals.
• Applied Support Vector Machine (SVM) multi-classification with selected features for online recognition.

Main Results:

• Demonstrated a robust indoor daily human activity recognition framework using a single Wi-Fi TP-AP pair.
• Achieved activity recognition that is insensitive to variations in location, orientation, and speed.
• Successfully segmented primitive action patterns from complex MIMO signals.

Conclusions:

• Wi-Fi CSI-based activity recognition is a viable and robust alternative/complement to vision-based systems.
• The proposed framework effectively recognizes complex activities composed of primitive actions.
• The system offers a practical solution for indoor activity monitoring in challenging environments.