Activity Recognition Invariant to Sensor Orientation with Wearable Motion Sensors.

Area of Science:

• Wearable computing
• Human activity recognition
• Sensor data processing

Background:

• Traditional activity recognition relies on fixed sensor placement.
• Real-world wearable sensor use involves variable positions and orientations.
• This variability degrades the performance of existing activity recognition systems.

Purpose of the Study:

• To develop wearable systems invariant to sensor orientation.
• To enable accurate activity recognition regardless of sensor placement.
• To enhance the practical applicability of wearable activity recognition.

Main Methods:

• Developed two novel transformations to eliminate absolute sensor orientation effects.
• Applied transformations to raw sensor data for pre-processing.
• Evaluated methodology with four state-of-the-art classifiers on five public datasets.

Main Results:

• Proposed transformations allow sensors to be worn at any orientation.
• Achieved activity recognition performance comparable to systems with fixed sensor orientations.
• Demonstrated robustness against incorrectly oriented sensors.

Conclusions:

• The proposed orientation-invariant techniques significantly improve wearable activity recognition robustness.
• These methods can be easily integrated into existing wearable systems.
• Enables more flexible and practical deployment of human activity recognition technology.