Single Accelerometer to Recognize Human Activities Using Neural Networks.

Area of Science:

• Robotics and Human-Computer Interaction
• Machine Learning and Artificial Intelligence
• Biomedical Engineering

Background:

• Exoskeletons reduce physical strain in daily activities but often lack versatility.
• Current assistance is activity-specific, limiting broader exoskeleton application.
• A robust human activity recognition (HAR) system is crucial for adaptable exoskeleton support.

Purpose of the Study:

• To develop and validate deep learning models for automatic human activity recognition.
• To enable adaptable exoskeleton assistance across diverse daily activities.
• To assess the efficacy of a single-sensor-based HAR system.

Main Methods:

• Two deep learning models were developed: a 1D convolutional neural network (CNN) and a hybrid CNN-LSTM model.
• Models were trained using data from a single three-axis accelerometer on the chest of ten subjects.
• Five activities (standing, level walking, incline walking, running, squatting) were classified.

Main Results:

• The CNN and hybrid models achieved high classification accuracies of 98.1% and 97.8% respectively in offline testing.
• Real-time validation with two subjects demonstrated high accuracies of 96.6% (CNN) and 97.2% (hybrid model).
• The findings indicate a single sensor can effectively distinguish human activities.

Conclusions:

• Deep learning models, particularly CNN and hybrid CNN-LSTM, show significant promise for HAR.
• A single accelerometer-based system can reliably recognize multiple human activities.
• This approach paves the way for more adaptive and versatile exoskeleton technologies.