Automatic fall detection using region-based convolutional neural network.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• Traditional fall detection methods rely on complex, often inadequate, feature extraction for machine learning.
• Existing approaches struggle to identify optimal classifiers and feature extraction techniques for reliable fall detection.

Purpose of the Study:

• To address limitations in feature extraction and classifier selection for fall detection using deep learning.
• To develop and evaluate a deep learning framework for accurate fall incident recognition.

Main Methods:

• Implemented a general framework utilizing a faster region-based convolutional neural network (R-CNN).
• Designed three custom deep learning architectures.
• Employed transfer learning with pre-trained networks (VGG-16, AlexNet) to enhance fall detection capabilities.

Main Results:

• Deep learning models, particularly those leveraging pre-trained networks, demonstrated superior performance in fall detection.
• VGG-16 achieved the highest accuracy, followed closely by AlexNet.
• Custom-designed networks also yielded impressive results, nearing the performance of pre-trained models.

Conclusions:

• Deep learning effectively automates feature learning, overcoming challenges in traditional fall detection systems.
• Transfer learning with pre-trained networks offers a significant advantage for fall detection tasks.
• The proposed framework provides a robust and accurate solution for recognizing fall incidents.