Reducing the Impact of Sensor Orientation Variability in Human Activity Recognition Using a Consistent Reference...

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Sensor orientation is crucial for accurate Human Activity Recognition (HAR) using tri-axial signals.
• Variations in sensor placement introduce significant errors, degrading HAR system performance.

Purpose of the Study:

• To develop and evaluate a novel preprocessing module to address sensor orientation variability in HAR.
• To improve the robustness and accuracy of HAR systems despite inconsistent sensor orientations.

Main Methods:

• A preprocessing module was designed to estimate a consistent reference system.
• Tri-axial sensor signals were transformed into this reference system.
• The method was validated across six diverse HAR datasets using subject-wise cross-validation.

Main Results:

• The proposed module effectively mitigated the negative impact of sensor orientation variability on HAR accuracy.
• Robust HAR performance was achieved even with sudden changes in sensor orientation.
• On the WISDM dataset, accuracy was recovered from 89.19% to 91.46%, matching the performance with consistent orientation.

Conclusions:

• The developed preprocessing module significantly enhances the reliability of HAR systems in real-world scenarios.
• This approach offers a practical solution for deploying HAR systems where sensor orientation cannot be strictly controlled.
• The method demonstrates substantial improvements in classification accuracy across multiple datasets.