An Unobtrusive Human Activity Recognition System Using Low Resolution Thermal Sensors, Machine and Deep Learning

Area of Science:

• Biomedical Engineering
• Computer Science

Background:

• The aging population necessitates advanced healthcare solutions for age-related health issues like injurious falls.
• Current fall detection methods often rely on obtrusive wearable devices, leading to poor user compliance.

Purpose of the Study:

• To develop and evaluate an unobtrusive human activity and fall detection system using infrared technology.
• To compare the performance of deep learning (DL) and machine learning (ML) approaches for activity recognition and location detection.
• To assess the effectiveness of a stereo vision system by fusing data from two infrared sensors.

Main Methods:

• Prototyped a system with two 24×32 thermal array sensors for unobtrusive monitoring.
• Collected data from healthy volunteers across ten different scenarios.
• Employed a supervised deep learning (DL) algorithm for activity classification and location detection, comparing it with machine learning (ML) methods.
• Implemented a stereo system by fusing data from two sensors.
• Performed binary classification to specifically detect critical activities like falling and lying on the floor.

Main Results:

• The DL-based stereo system achieved high accuracy in activity recognition (97.6%) and location detection (97.3%).
• Binary classification for critical activities (falling/lying) reached 97.9% accuracy.
• The stereo system demonstrated superior performance compared to a single sensor setup.
• Overall F1-scores for activity recognition, location detection, and binary classification were 0.935, 0.927, and 0.945, respectively.

Conclusions:

• The proposed unobtrusive infrared monitoring system accurately recognizes human activities, detects locations, and identifies critical events like falls.
• This system presents a viable, non-wearable solution for enhanced safety and monitoring in healthcare settings, particularly for the elderly.