Fall Detector Adapted to Nursing Home Needs through an Optical-Flow based CNN.

Area of Science:

• Gerontology
• Computer Science
• Medical Imaging

Background:

• Fall detection in elderly care is crucial but challenging.
• Vision-based systems offer advantages over sensor-based methods for residents with cognitive impairments.
• Existing solutions require optimization for real-world nursing home deployment.

Purpose of the Study:

• To develop and evaluate a Deep Learning-based vision fall detection system for nursing homes.
• To incorporate medical team requirements into the system's tuning and evaluation.
• To address the challenges of fall detection in specialized elderly care environments.

Main Methods:

• Utilized a Convolutional Neural Network (CNN) for fall detection.
• Trained the CNN to optimize a sensitivity-based metric, considering medical requirements.
• Developed a custom metric and decision-making process emphasizing the temporal aspect of falls.

Main Results:

• The proposed system achieved an 86.2% fall detection rate.
• The system generated an average of 11.6% false alarms across tested databases.
• Results underscore the significance of temporal data in accurate fall detection.

Conclusions:

• The developed Deep Learning model effectively detects falls in elderly care settings.
• Tailoring the CNN and decision process to medical needs enhances clinical relevance.
• Vision-based fall detection with temporal analysis shows promise for nursing home safety.