Real-Time Sensor-Based Human Activity Recognition for eFitness and eHealth Platforms.

Area of Science:

• Computer Science
• Biomedical Engineering
• Sports Science

Background:

• Human Activity Recognition (HAR) is crucial for applications in healthcare, security, and entertainment.
• Accurate exercise tracking and repetition counting are essential for athletic training and monitoring.
• Existing HAR methods often struggle with real-time performance and robustness against background activities.

Purpose of the Study:

• To develop a real-time system for recognizing 12 types of exercises and counting repetitions during athletic workouts.
• To improve the accuracy and robustness of HAR models using deep neural networks and motion sensor data.
• To investigate the impact of design requirements and data collection protocols on HAR system performance.

Main Methods:

• Utilized a deep neural network model fed by data from a 9-axis Inertial Measurement Unit (IMU) sensor placed on the chest.
• Implemented an encoder architecture pretrained with contrastive learning for improved model quality.
• Designed the system for deployment on mobile platforms (iOS, Android) for accessibility.

Main Results:

• The proposed approach significantly enhances model accuracy (F1 score, MAPE) compared to end-to-end training.
• Demonstrated improved robustness against background activity, indicated by a lower false-positive rate.
• The system achieves real-time recognition and repetition counting for 12 distinct exercises.

Conclusions:

• The developed HAR system offers a high-quality, robust, and accurate solution for exercise recognition and counting.
• Pretraining with contrastive learning is an effective strategy for improving deep learning models in HAR.
• The public release of the AIDLAB-HAR dataset will foster further research in mobile HAR systems.