Sensor-Based Human Activity Recognition Using Adaptive Class Hierarchy.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Sensor-based human activity recognition commonly employs Convolutional Neural Networks (CNNs).
• Traditional CNN models treat activity classes independently, neglecting inherent hierarchical relationships.
• Recognizing hierarchical structures in activities can enhance model performance.

Purpose of the Study:

• To propose a novel Branch CNN (B-CNN) model that automatically constructs class hierarchies for activity recognition.
• To address the limitations of manual hierarchy design in B-CNNs, especially with large datasets or limited prior knowledge.

Main Methods:

• Developed a class hierarchy-adaptive B-CNN that integrates automatic hierarchy construction from training data.
• The proposed method enables effective B-CNN training without requiring pre-defined or manually crafted class hierarchies.
• Evaluated the approach on multiple benchmark activity recognition datasets.

Main Results:

• The proposed method significantly outperformed standard CNN models that do not leverage class hierarchies.
• Achieved performance comparable to B-CNN models that utilize human-defined class hierarchies.
• Demonstrated the effectiveness of automatically generated hierarchies in sensor-based activity recognition.

Conclusions:

• Automatic class hierarchy construction is a viable and effective strategy for enhancing CNN-based activity recognition.
• The class hierarchy-adaptive B-CNN offers a robust alternative to manual hierarchy design, improving recognition accuracy.
• This approach advances the field by enabling more sophisticated modeling of complex activity relationships.