Automatic Annotation for Human Activity Recognition in Free Living Using a Smartphone.

Area of Science:

• Computer Science
• Machine Learning
• Human-Computer Interaction

Background:

• Data annotation is a significant bottleneck in developing Human Activity Recognition (HAR) systems, particularly for personalized and online applications.
• Supervised machine learning (ML) methods require large, labeled datasets, which are difficult and time-consuming to acquire, especially for user-specific HAR data.
• Addressing inter-person variability in smartphone-based HAR necessitates specialized, labeled datasets.

Purpose of the Study:

• To present an automated method for labeling datasets collected under free-living conditions using smartphones.
• To evaluate the robustness of common supervised classification algorithms when faced with noisy data labels.
• To facilitate the creation of larger, more diverse datasets for HAR research.

Main Methods:

• Development of an automatic labeling technique for smartphone-based data collection in natural environments.
• Investigation of the performance of various supervised ML classifiers (Neural Networks, Random Forests, Nearest Centroid, Multi-Class SVM) on automatically labeled and noisy datasets.
• Validation using a 38-day manually labeled dataset collected in free-living conditions.

Main Results:

• The automatic labeling method achieved an average precision rate of 80-85% compared to manual labeling.
• Supervised models trained on automatically generated labels attained an 84% f-score (Neural Networks, Random Forests).
• Label noise significantly impacted performance, reducing the f-score to 64-74% for certain classifiers (Nearest Centroid, Multi-Class SVM).

Conclusions:

• Automated data labeling is a viable strategy to accelerate HAR dataset creation in real-world settings.
• While automated labels yield strong ML performance, classifier robustness to label noise is crucial for reliable HAR systems.
• Further research should focus on enhancing noise-resilient algorithms for practical, personalized HAR applications.