Pre-Impact Fall Detection Using 3D Convolutional Neural Network.

Area of Science:

• Biomedical Engineering
• Computer Science
• Rehabilitation Medicine

Background:

• Early fall detection is crucial for patient safety during gait rehabilitation.
• Existing methods often lack accuracy or timeliness, especially with limited data.

Purpose of the Study:

• To develop and evaluate a pre-impact fall detection system using a 3D convolutional neural network (CNN).
• To address the challenge of limited training data in gait rehabilitation settings.

Main Methods:

• A 3D CNN was pre-trained on general walking and fall data.
• The network was fine-tuned with trainee-specific data for personalized fall detection.
• A temporal sliding window approach was used with RGB video input.

Main Results:

• The system achieved 100% detection accuracy within 0.5 seconds of fall onset.
• The training strategy effectively mitigated generalization issues with limited data.
• The approach demonstrated efficiency, accuracy, and practicality.

Conclusions:

• This is the first pre-impact fall detection method using a 3D CNN with RGB images.
• The proposed method offers a robust solution for intelligent fall detection in gait rehabilitation training.