Ensemble of RNN Classifiers for Activity Detection Using a Smartphone and Supporting Nodes

Area of Science:

• Human-Computer Interaction
• Wearable Technology
• Biomedical Engineering

Background:

• Mobile devices accurately recognize basic daily activities.
• Existing methods struggle with complex training routines (e.g., squats, jumps).
• Increased activity sets reduce recognition accuracy.

Purpose of the Study:

• Propose a personal area network model to enhance human activity recognition accuracy.
• Utilize a smartphone as the main node with supporting sensor nodes.
• Improve recognition of diverse and complex physical activities.

Main Methods:

• Implemented a personal area sensor network with a smartphone and body-attached sensors.
• Employed recurrent neural networks (RNNs) on sensor nodes for local activity categorization.
• Used a weighted voting procedure on the main node for final recognition.
• Optimized sensor node reporting to conserve energy.

Main Results:

• Achieved higher recognition accuracy for a set of eight activities compared to existing methods.
• Evaluated performance with sensors on waist, chest, leg, and arm.
• Determined optimal sensor node configuration for maximum accuracy and reduced transmissions.

Conclusions:

• The proposed personal area sensor network significantly improves human activity recognition accuracy.
• Distributed processing on sensor nodes combined with central fusion is effective.
• Energy-saving strategies enhance network lifetime and practicality.