Impact of Gyroscope Integration, Sensor Placement, and Activity Granularity on Human Activity Recognition Performance

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Human Activity Recognition

Background:

• Accurate human activity recognition (HAR) using wearable sensors is crucial for population health studies.
• Optimizing sensor configuration and classification strategies is essential for reliable data collection.

Purpose of the Study:

• To evaluate the impact of sensor configuration, body location, classification granularity, and model choice on HAR.
• To identify optimal strategies for wearable sensing protocols in large-scale cohort studies like the Spanish IMPaCT cohort.

Main Methods:

• Collected data from 85 participants using thigh-, wrist-, and hip-mounted inertial measurement units (IMUs).
• Analyzed signals using convolutional neural networks, Random Forest, and XGBoost classifiers with overlapping 10-s windows.
• Defined two classification targets: fine-grained (15 activities) and coarse-grained (4 MET-based intensity levels).

Main Results:

• Classification granularity was the primary performance determinant, with intensity-level classification outperforming fine-grained activity recognition.
• Wrist-mounted sensors yielded the highest F1-scores, and gyroscope data consistently improved performance.
• Sensor configuration, body location, and model type significantly influenced HAR outcomes.

Conclusions:

• Coarse-grained intensity classification offers higher accuracy than fine-grained activity recognition for HAR.
• Optimal sensor placement (e.g., wrist) and the inclusion of gyroscope data are key for enhancing HAR.
• Findings provide practical guidance for designing wearable sensing protocols for population-based studies and real-world applications.