Real-time activity and fall detection using transformer-based deep learning models for elderly care applications

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Computer Science

Background:

• Existing activity recognition and fall detection methods struggle with accuracy and real-time performance.
• Wearable sensor data analysis is crucial for monitoring daily living activities and detecting falls.

Purpose of the Study:

• To develop a transformer-based deep learning model for accurate and real-time activity recognition and fall detection.
• To address the limitations of current systems in terms of accuracy and real-time applicability.

Main Methods:

• Utilized a transformer encoder with a self-attention mechanism to process wearable sensor data (accelerometer, gyroscope, orientation) via sliding window segmentation.
• Trained and evaluated the model on the extensive MobiAct dataset, comprising over 14 million records from 66 participants across 16 activities, including various fall types.

Main Results:

• Achieved over 98% accuracy, with excellent precision and recall for complex fall categories like forward-lying and sideward-lying.
• Demonstrated superior performance compared to Convolutional Neural Networks-Long Short-Term Memory (CNN-LSTM) and Temporal Convolutional Networks across classification metrics and training stability.

Conclusions:

• Transformer models effectively capture complex temporal dependencies, mitigating misclassification and false positives in activity recognition and fall detection.
• The developed transformer-based system offers efficient real-time deployment and reliable solutions for elderly care and fall prevention.