Activity Recognition Invariant to Wearable Sensor Unit Orientation Using Differential Rotational Transformations...

Area of Science:

• Biomedical Engineering
• Human-Computer Interaction
• Signal Processing

Background:

• Wearable motion sensors are crucial for activity recognition.
• Existing algorithms often fail when sensors have variable orientations on the body.
• This limitation hinders the reliability of wearable health and activity monitoring systems.

Purpose of the Study:

• To develop a novel method for transforming motion sensor data to be invariant to sensor unit orientation.
• To enhance the robustness of activity recognition algorithms against sensor placement variations.
• To improve the accuracy and reliability of wearable sensor systems.

Main Methods:

• Estimating sensor unit orientation and transforming data to the Earth frame.
• Representing sensor rotations between time samples using quaternions in the Earth frame.
• Integrating the orientation-invariant transformation into the pre-processing stage of activity recognition.

Main Results:

• The proposed method achieved an average accuracy drop of only 4.7% with incorrectly oriented sensors, compared to 31.8% for standard systems.
• Outperformed existing orientation-robust methods, which showed accuracy degradation between 8.4% and 18.8%.
• Demonstrated superior performance in classifying stationary activities where other methods notably failed, with a 2.1% accuracy drop for non-stationary activities.

Conclusions:

• The proposed orientation-invariant method significantly enhances the robustness of wearable motion sensor systems.
• This approach is broadly applicable to various wearable devices for reliable activity recognition regardless of sensor orientation.
• The method effectively addresses a key limitation in current wearable sensor technology.