Hierarchical multi-view aggregation network for sensor-based human activity recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Sensor-based human activity recognition (HAR) is crucial for monitoring physical activities.
• Existing deep learning methods often rely on black-box feature extraction from raw sensor data.
• A need exists for more interpretable and robust HAR models.

Purpose of the Study:

• To propose a novel hierarchical multi-view aggregation network for sensor-based HAR.
• To enhance HAR accuracy by effectively integrating multi-view sensor features.
• To develop a model that captures complex relationships across sensor features, positions, and modalities.

Main Methods:

• Constructing multi-view feature spaces (white-box and black-box) for individual sensors.
• Developing a hierarchical aggregation network to learn unified representations from multi-view features.
• Designing three aggregation modules for feature, position, and modality level integration.
• Utilizing non-local operations and attention mechanisms for feature correlation and relationship modeling.

Main Results:

• The proposed method achieved superior accuracy on 12 human activity benchmark datasets.
• Demonstrated effective integration of diverse sensor data views.
• Outperformed state-of-the-art approaches in sensor-based HAR.

Conclusions:

• The hierarchical multi-view aggregation network offers a powerful approach to sensor-based HAR.
• Integrating white-box and black-box features enhances model performance.
• The method's ability to capture cross-level relationships is key to its success.