TCS-Fall: Cross-individual fall detection system based on channel state information and time-continuous stack method.

Area of Science:

• Elderly care technology
• Wireless sensing
• Machine learning for health

Background:

• Falls present a significant health risk for older adults, especially those living alone.
• WiFi-based fall detection using Channel State Information (CSI) offers a non-intrusive and privacy-preserving solution.
• A key challenge is optimizing the performance of CSI-based fall detection across different individuals.

Purpose of the Study:

• To develop a resilient, real-time fall detection system (TCS-Fall) that works across individuals using CSI.
• To enable continuous monitoring for accurate and prompt fall detection over extended periods.
• To address the challenge of cross-individual performance in CSI-based fall detection.

Main Methods:

• Collected extensive CSI data from 20 volunteers, including 1800 falls and 2400 daily activities.
• Utilized grouped coefficient of variation of CSI amplitudes as input features for a convolutional neural network classifier.
• Developed a user-friendly CSI data collection and detection tool with PyQT and optimized processing with Numba for real-time performance.

Main Results:

• The TCS-Fall method achieved excellent cross-individual fall detection performance (AUC 0.999, no error 0.955, correct warning 0.975) with data from only two volunteers.
• Performance improved to 1.00 AUC with data from 10 volunteers.
• Optimized data processing resulted in over a 20x speedup, with detection within 100 ms using the PyQT tool.

Conclusions:

• The TCS-Fall method provides effective real-time, cross-individual fall detection using WiFi CSI.
• The system promises swift alerts and timely assistance for the elderly.
• Optimized data processing significantly enhances system speed, highlighting its potential for real-world application.