Exploring human activity recognition using feature level fusion of inertial and electromyography data

Area of Science:

• Biomedical Engineering
• Human Activity Recognition
• Wearable Technology

Background:

• Wearable sensors are crucial for objective human activity recognition (HAR).
• Multi-modal sensing in wearables allows for combining different data types like inertial and surface electromyography (sEMG).
• Optimal methods for combining these modalities and processing sEMG data for HAR remain under investigation.

Purpose of the Study:

• To investigate the efficient combination of inertial and sEMG data using machine learning for improved HAR.
• To evaluate the impact of different sEMG post-processing techniques on HAR accuracy.
• To recognize four basic daily life activities: walking, standing, stair ascent, and stair descent.

Main Methods:

• Developed a novel feature vector incorporating domain knowledge from mobility studies.
• Employed a feature-level data fusion approach to integrate inertial and sEMG data.
• Utilized Support Vector Machine (SVM) and k-Nearest Neighbors (kNN) classifiers with 5-fold cross-validation.

Main Results:

• The fusion of inertial data with sEMG increased overall HAR accuracy by 3.5% (SVM) and 6.3% (kNN).
• Extracting features from linear envelopes of sEMG signals, compared to bandpass filtered signals, improved HAR accuracy for both classifiers.
• Post-processing sEMG signals demonstrated a significant positive impact on multimodal HAR performance.

Conclusions:

• Combining inertial and sEMG data via feature-level fusion enhances HAR accuracy.
• sEMG signal post-processing, specifically using linear envelopes, is beneficial for improving HAR.
• These findings suggest that optimized multimodal wearable sensor data processing can advance HAR applications.