Accurate and Real-Time Hierarchical Ensemble Network for Activity Classification in Construction Worker

Area of Science:

• Robotics
• Human-Computer Interaction
• Machine Learning

Background:

• Accurate, real-time locomotion classification is vital for exoskeletons assisting construction workers.
• Current algorithms struggle with accuracy, real-time performance, and adaptability across different exoskeleton types.

Purpose of the Study:

• To develop a novel hierarchical ensemble network for classifying 11 lower limb activities in construction workers.
• To create adaptable solutions using various Inertial Measurement Unit (IMU) sensor placements for diverse wearable devices.

Main Methods:

• Developed a hierarchical ensemble network framework for activity classification.
• Created 6 configurations with IMU sensors on different body segments.
• Validated results using leave-one-out cross-validation with 10 subjects.

Main Results:

• The proposed algorithm demonstrated improved accuracy, precision, recall, and F1-score compared to a baseline ANN.
• Achieved significant reductions in model parameters (71.86%) and inference time (47.85%).
• Validated effectiveness across 6 different sensor configurations.

Conclusions:

• The hierarchical ensemble network provides accurate and real-time classification of multiple construction worker activities.
• Multiple sensor configurations enhance practicality and applicability for various assistive devices.
• The developed system offers high performance and deployable solutions for exoskeleton control.