SVM versus MAP on accelerometer data to distinguish among locomotor activities executed at different speeds

Area of Science:

• Biomechanics
• Machine Learning
• Wearable Technology

Background:

• Accurate classification of human locomotor activities is crucial for applications in healthcare and sports science.
• Existing methods often struggle with naturalistic, varied movement patterns.
• Inertial sensors offer a promising avenue for unobtrusive activity monitoring.

Purpose of the Study:

• To compare the efficacy of a Maximum a Posteriori (MAP) Bayes' classification scheme against a Support Vector Machine (SVM) for classifying locomotor activities.
• To evaluate these methods using features extracted from accelerometer data.
• To assess performance in a natural indoor-outdoor environment.

Main Methods:

• 16 features were extracted from accelerometer data in the time and frequency domains.
• 2D Sammon's mapping was employed for dimension reduction.
• An Artificial Neural Network (ANN) mimicked Sammon's mapping.
• A Bayes' classifier and an SVM were trained and tested on the reduced feature set.

Main Results:

• The Bayes' approach achieved higher accuracy than SVM on both training (91.4% vs. 90.7%) and testing datasets (84.2% vs. 76.0%).
• Both methods utilized 2D projections of 16 extracted features for classification.
• Activities included level walking, stair climbing, and stair descending.

Conclusions:

• The proposed Bayes' classification scheme is more suitable than the SVM for distinguishing between different locomotor activities.
• Accelerometer-based feature extraction and dimension reduction are effective for activity recognition.
• The findings support the use of Bayes' classifiers in wearable sensor-based human activity monitoring.