An Information Gain-Based Model and an Attention-Based RNN for Wearable Human Activity Recognition

Area of Science:

• Computer Science
• Biomedical Engineering
• Machine Learning

Background:

• Wearable sensor-based human activity recognition (HAR) is crucial for health monitoring and human-computer interaction.
• Current HAR systems lack a unified human activity model, leading to inconsistent sensor placement and quantity, hindering widespread adoption.
• Standardized models are needed to improve the efficiency and applicability of wearable HAR.

Purpose of the Study:

• To establish an information gain-based human activity model for guiding sensor deployment in wearable HAR systems.
• To design and evaluate an attention-based recurrent neural network (Attention-RNN) for enhanced human activity recognition.
• To demonstrate the potential for reducing sensor count while maintaining classification accuracy.

Main Methods:

• Developed an information gain-based model to determine optimal sensor locations and quantities.
• Designed an Attention-RNN architecture, integrating Bidirectional Long Short-Term Memory (BiLSTM) with an attention mechanism.
• Validated the model and Attention-RNN on the UCI Opportunity Challenge dataset.

Main Results:

• The proposed human activity model effectively guides sensor deployment, enabling reduced sensor numbers without compromising classification performance.
• The Attention-RNN achieved high F1 scores: 0.898 for Modes of Locomotion (ML) and 0.911 for Gesture Recognition (GR).
• Experimental results confirm the efficacy of the integrated approach for robust HAR.

Conclusions:

• The information gain-based model provides a framework for optimizing sensor configuration in wearable HAR.
• The Attention-RNN demonstrates superior performance in recognizing human activities using wearable sensor data.
• This research contributes to more efficient and effective wearable HAR systems.