Human activity recognition algorithms for manual material handling activities.

Area of Science:

• Biomedical Engineering
• Ergonomics
• Machine Learning

Background:

• Human Activity Recognition (HAR) using Inertial Measurement Units (IMUs) is increasingly important for ergonomic evaluation and exoskeleton technology.
• Current research lacks comprehensive studies on identifying diverse lifting styles, necessitating optimized datasets and classification algorithms.

Purpose of the Study:

• To investigate the impact of sensor placement, quantity, time window, classifier complexity, and IMU data types on lifting style classification.
• To identify optimal parameters for accurate lifting style detection, crucial for exoskeleton-assisted strategies and injury prevention.

Main Methods:

• Analysis of feedforward neural networks, 1-D convolutional neural networks, and recurrent neural networks for time series classification.
• Evaluation of various sensor configurations, time window durations, and IMU data types.

Main Results:

• Accurate lifting style detection is significantly influenced by the number of sensors used and the duration of the time window.
• Classifier architectures capable of leveraging temporal data dependencies are essential for distinguishing subtle kinematic differences in lifting styles.

Conclusions:

• Optimal human activity recognition for lifting styles necessitates multiple sensors and extended time windows.
• Advanced neural network architectures are vital for accurate classification, particularly for embedded systems like occupational exoskeletons, to mitigate injury risks.