Towards Human Activity Recognition: A Hierarchical Feature Selection Framework.

Area of Science:

• Computer Science
• Biomedical Engineering
• Human-Computer Interaction

Background:

• Human activity recognition (HAR) using wearable sensors is challenging due to the complexity of physical activities.
• Existing methods often struggle with high-dimensional feature spaces and achieving accurate recognition.

Purpose of the Study:

• To propose a novel hierarchical activity recognition framework to improve performance.
• To introduce and evaluate two distinct feature selection methods for optimizing feature spaces within the hierarchical structure.

Main Methods:

• A hierarchical tree structure is employed to organize predefined activities, with internal nodes representing activity groups and leaf nodes representing specific activities.
• Two feature selection methods are integrated to optimize the feature space at each node of the hierarchy.
• Classifiers are trained to distinguish activity groups and predict activity labels in a top-down manner.

Main Results:

• The proposed framework and feature selection methods effectively reduce feature space dimensionality.
• Experimental results demonstrate a significant enhancement in overall human activity recognition accuracy.
• Activity-specific feature subsets generally yield superior performance compared to a common feature subset for all activities.

Conclusions:

• The hierarchical framework combined with optimized feature selection offers a robust approach to human activity recognition.
• The findings suggest that tailoring feature selection to specific activity levels within a hierarchy is beneficial for improved accuracy.
• This work contributes to more accurate and efficient activity recognition systems for wearable sensor applications.