Privacy-Preserved Fall Detection Method with Three-Dimensional Convolutional Neural Network Using Low-Resolution...

Area of Science:

• Gerontology
• Computer Science
• Biomedical Engineering

Background:

• Increasing elderly populations in healthcare facilities strain staff resources.
• Existing human motion detection systems face limitations like user inconvenience, privacy concerns, and environmental dependencies (e.g., lighting).
• Real-time monitoring is crucial for promptly addressing emergencies like falls among vulnerable individuals.

Purpose of the Study:

• To develop a privacy-preserving human motion detection system for elderly care settings.
• To overcome the limitations of existing systems using infrared sensor technology.
• To accurately detect various human motions, with a specific focus on falls.

Main Methods:

• Development of a human motion detection system utilizing a low-resolution infrared array sensor.
• Implementation of a three-dimensional convolutional neural network (3D CNN) for motion analysis.
• Experimental validation with 16 participants, including cross-validation for fall detection.

Main Results:

• The proposed system achieved 98.8% accuracy and 94.9% F1-measure in fall detection.
• Performance metrics were superior to those of a long short-term memory (LSTM) network by 1% and 3.6% respectively.
• The system demonstrated the capability to detect eight distinct types of human motion.

Conclusions:

• The developed infrared sensor-based system offers a viable solution for real-time elderly safety monitoring.
• The system effectively addresses privacy and lighting condition limitations inherent in other detection methods.
• The high accuracy and feasibility for practical application suggest significant potential in hospitals and nursing homes.