Classifier Personalization for Activity Recognition Using Wrist Accelerometers.

Area of Science:

• Exercise Science
• Biomedical Engineering
• Human-Computer Interaction

Background:

• Accelerometer-based activity recognition faces challenges due to significant intersubject variability, impacting classification accuracy and limiting generalizability to new users.
• Existing methods often struggle to adapt to individual differences in movement patterns, hindering reliable real-time activity tracking.

Purpose of the Study:

• To develop and evaluate a personalized approach for activity recognition using accelerometer data.
• To improve the accuracy and reliability of classifying activities like ambulation, cycling, sedentary behavior, and other activities for individual users.

Main Methods:

• Extended a support vector machine (SVM) based activity classification method to estimate classification uncertainty.
• Utilized uncertainty estimation to prompt user-driven data labeling, which was then used to update the personalized classifier.
• Evaluated the method using two datasets (adults and youth) with multiple activity types, employing leave-one-subject-out and leave-one-group-out cross-validation.

Main Results:

• The personalized classification method demonstrated an average improvement in overall recognition accuracy of up to 11%.
• Observed significant person-specific accuracy improvements, ranging from -2% to +36%.
• The approach proved effective across diverse datasets and activity types.

Conclusions:

• Personalization of accelerometer-based activity recognition rules significantly enhances classification accuracy for individual users.
• The proposed method is suitable for online implementation, supporting real-time activity recognition systems.
• This approach addresses intersubject variability, paving the way for more reliable and user-specific activity tracking applications.