Fitness activity classification by using multiclass support vector machines on head-worn sensors

Area of Science:

• Wearable technology
• Human activity recognition
• Biomedical engineering

Background:

• Wearable devices offer personalized fitness insights.
• Optical head-mounted displays (OHMD) are emerging for enhanced user interaction.
• Accurate fitness activity classification is crucial for performance metrics.

Purpose of the Study:

• To introduce a novel method for fitness activity classification using head-worn sensors.
• To compare the efficacy of head-worn sensors against other body locations.
• To determine optimal sensor placement for accurate activity recognition.

Main Methods:

• Utilized accelerometer, barometric pressure sensor, and GPS from head-worn devices.
• Employed multiclass Support Vector Machine (SVM) for classification.
• Evaluated sensor performance across various body locations, focusing on head-worn configurations.

Main Results:

• Achieved an average F-score of 96.66% for classifying standing, walking, running, stair climbing, and cycling.
• Identified ankle plus another upper body location as the best sensor combination.
• Found no significant improvement with three or more sensors compared to the best two-sensor setups.

Conclusions:

• Head-worn sensors, particularly with SVM, are highly effective for fitness activity classification.
• Optimal sensor placement significantly impacts classification accuracy, with combined ankle and upper body locations being superior.
• Multi-sensor fusion beyond two locations does not substantially enhance performance for this task.