Classification of periodic activities using the Wasserstein distance

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Accelerometer data analysis is crucial for understanding human activity recognition.
• Existing methods may struggle with variations in activity speed and incline.
• Nonparametric classification offers a flexible approach to pattern recognition.

Purpose of the Study:

• To introduce a novel nonparametric classification technique utilizing Wasserstein distance.
• To apply this method for classifying periodic activities (walking, biking, running) from accelerometer data.
• To evaluate the system's performance on a diverse subject corpus.

Main Methods:

• Utilized Wasserstein distance for nonparametric classification of time-frequency representations of accelerometer data.
• Focused on the global structure of frequency patterns rather than specific peak locations.
• Integrated supervised learning techniques with the Wasserstein distance approach.

Main Results:

• The Wasserstein distance method demonstrated robustness in detecting activities irrespective of speed or incline.
• The proposed classification scheme achieved performance comparable to more complex systems.
• Successfully classified three distinct periodic activities: walking, biking, and running.

Conclusions:

• Wasserstein distance provides an effective and robust method for nonparametric classification of accelerometer data.
• This approach offers a valuable tool for biomedical applications in activity recognition.
• The technique's insensitivity to speed and incline enhances its practical utility.