Detecting falls and estimation of daily habits with depth images using machine learning algorithms

Area of Science:

• Gerontology
• Computer Science
• Biomedical Engineering

Background:

• Falls are a significant risk for elderly individuals, necessitating effective detection methods.
• Existing fall detection systems have limitations in accuracy and real-world applicability.

Purpose of the Study:

• To develop and compare supervised learning methods for fall detection using depth image parameters.
• To evaluate the performance of Decision Tree, K-Nearest Neighbors (K-NN), and Random Forests (RF) algorithms.
• To propose a method for estimating daily habits through long-term monitoring.

Main Methods:

• Extracted parameters from depth images captured in a nursing home environment.
• Compared three supervised learning algorithms: Decision Tree, K-NN, and Random Forests.
• Tested methods on a 43-day database of depth images.
• Conducted a 37-day follow-up for daily habit estimation.

Main Results:

• Random Forests (RF) demonstrated superior performance in fall detection.
• RF achieved 93% sensitivity and 100% specificity, particularly when focused on the area around the bed.
• The study successfully proposed a method for estimating daily habits.

Conclusions:

• Supervised learning, specifically Random Forests, offers a highly accurate approach to fall detection in elderly care.
• Depth image analysis is a viable technique for developing effective fall detection systems.
• Long-term monitoring using this method can provide insights into daily habits of elderly individuals.