Fall detection among elderly persons using FallCNN and transfer learning models

Area of Science:

• Gerontology
• Computer Science
• Biomedical Engineering

Background:

• Falls are a leading cause of injury and death in older adults globally.
• Existing fall detection methods lack efficient artificial intelligence (AI) strategies.
• Fall Detection among Elderly Persons (FDEP) offers a potential cost-effective solution.

Purpose of the Study:

• To develop an efficient AI-based fall detection system for early intervention.
• To create a novel signal-based image dataset (SimgFall) for fall detection.
• To design and evaluate a deep Convolutional Neural Network (CNN) architecture, FallCNN, for improved fall detection accuracy.

Main Methods:

• Generated the SimgFall dataset comprising 1992 signal-based images from accelerometer/gyroscope data.
• Developed FallCNN, a novel CNN architecture with depth-wise convolution and varying dilation rates.
• Trained and evaluated multiple FallCNN variants (FallCNN_1 to FallCNN_4) on the SimgFall dataset.

Main Results:

• The FallCNN model achieved a highest classification accuracy of 98% with a loss of 0.0833.
• Architectural enhancements in FallCNN variants led to cumulative performance improvements, from 94% to 98% accuracy.
• The SimgFall dataset effectively supported the training and evaluation of deep learning models for fall detection.

Conclusions:

• The FallCNN model demonstrates high efficacy in early fall detection for elderly persons.
• The SimgFall dataset provides a valuable resource for developing advanced AI-driven fall detection systems.
• This AI strategy has the potential to significantly accelerate medical assistance for falls in older adults.