A Novel Deep Neural Network Method for HAR-Based Team Training Using Body-Worn Inertial Sensors.

Area of Science:

• Computer Science
• Biomedical Engineering
• Machine Learning

Background:

• Human Activity Recognition (HAR) is a challenging problem, especially with vision-based systems facing issues like low illumination and occlusion.
• Wearable inertial sensors offer a privacy-preserving alternative, adept at overcoming vision-based system limitations.

Purpose of the Study:

• To propose a novel approach for indistinguishable human activity recognition using wearable sensors.
• To develop a robust and effective HAR model that overcomes limitations of existing methods.

Main Methods:

• A multistage deep neural network framework was developed, interpreting data from accelerometers, gyroscopes, and magnetometers.
• Variational Autoencoders (VAE) were used for crucial information extraction from Inertial Measurement Unit (IMU) data.
• Generative Adversarial Networks (GANs) were employed to generate realistic human activity data.
• Transfer learning was applied to enhance model performance in the target domain.

Main Results:

• The proposed framework effectively extracts key information from raw sensor data.
• The integration of GANs improved the generation of realistic human activity data.
• Transfer learning significantly enhanced the model's performance for human activity recognition.

Conclusions:

• The developed multistage deep neural network framework provides a robust and effective solution for human activity recognition using wearable sensors.
• This approach successfully addresses challenges faced by vision-based HAR systems, offering a privacy-conscious alternative.