A User-Adaptive Algorithm for Activity Recognition Based on K-Means Clustering, Local Outlier Factor, and...

Area of Science:

• Computer Science
• Biomedical Engineering
• Human-Computer Interaction

Background:

• Mobile activity recognition is crucial for human-centric applications, but user variability in sensor data distribution degrades classifier performance.
• Personalized models are needed to overcome the challenge of transferring activity recognition models between users.

Purpose of the Study:

• To develop a personalized, user-adaptive algorithm for recognizing four categories of human activities (light, moderate, vigorous intensity, and falls).
• To address the performance degradation of activity recognition classifiers when applied to new users due to differing inertial sensor data distributions.

Main Methods:

• A user-adaptive algorithm integrating K-Means clustering, local outlier factor (LOF), and multivariate Gaussian distribution (MGD) was proposed.
• An improved K-Means algorithm with a novel initialization method was designed for automatic clustering and annotation of user-specific activity data.
• A method for quantifying sample informativeness was used to select the most valuable data for activity recognition model adaptation.

Main Results:

• The proposed algorithm demonstrated effective adaptation to new users.
• The personalized classifier achieved good recognition performance across different users.

Conclusions:

• The developed user-adaptive algorithm successfully overcomes the challenge of user variability in mobile activity recognition.
• This approach enhances the robustness and applicability of activity recognition systems in real-world scenarios, particularly for personalized applications.