Sensor Data Acquisition and Multimodal Sensor Fusion for Human Activity Recognition Using Deep Learning.

Area of Science:

• Biomedical Engineering
• Computer Science
• Wearable Technology

Background:

• Human Activity Recognition (HAR) systems rely on accurate data from body-worn sensors.
• Optimizing sensor positioning and data acquisition is crucial for HAR system efficiency and effectiveness.

Purpose of the Study:

• To systematically investigate optimal on-body sensor positioning and data acquisition for HAR systems.
• To develop and evaluate a deep learning framework for HAR using multi-sensor data.
• To explore sensor fusion techniques for improved activity classification performance.

Main Methods:

• A testbed with eight Inertial Measurement Units (IMU) sensors and an Android mobile device was utilized for data collection.
• A Long Short-Term Memory (LSTM) network framework was developed for training deep learning models.
• Experiments were conducted in both real-world and controlled environments.

Main Results:

• Activity data from four sensors (wrists, right ankle, waist) at a low sampling rate (10 Hz) proved sufficient for recognizing Activities of Daily Living (ADLs).
• A two-level ensemble model demonstrated that classifier-level sensor fusion enhances classification performance.
• Custom weights for multimodal sensor fusion were elaborated based on individual activity characteristics.

Conclusions:

• Optimal sensor selection and placement can significantly reduce data requirements for effective HAR.
• Deep learning models, particularly LSTMs, combined with sensor fusion, offer robust solutions for HAR.
• The findings provide practical guidelines for designing efficient and accurate HAR systems.