Deep Learning Multi-Class Approach for Human Fall Detection Based on Doppler Signatures.

Area of Science:

• Engineering
• Computer Science
• Gerontology

Background:

• Falling events pose a significant global health risk, particularly for the elderly, leading to severe physical and psychological consequences.
• Existing fall detection methods often rely on wearable sensors or dedicated hardware, limiting their practicality and widespread adoption.

Purpose of the Study:

• To develop and evaluate a non-wearable human fall detection system using radio frequency (RF) sensing.
• To investigate the feasibility of integrating fall detection into existing wireless communication infrastructure.

Main Methods:

• A sensing platform utilizing a continuous wave (CW) radio-frequency (RF) probe signal was developed.
• Changes in the probe signal's phase, amplitude, and frequency, imprinted with Doppler signatures by human movement, were analyzed.
• Two deep learning models, Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN), were trained to classify falling events.

Main Results:

• The developed RF sensing system achieved a fall detection accuracy of 92.1% for both LSTM and CNN models.
• The system successfully differentiated falling events from other common indoor activities.

Conclusions:

• The study demonstrates the viability of using a CW RF probe signal for non-contact human fall detection.
• This radio sensing approach offers a promising, easily integrable solution for elderly fall monitoring, potentially enhancing safety and reducing healthcare burdens.