Wearable sensors based on artificial intelligence models for human activity recognition.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Human motion detection is crucial for medicine, healthcare, and physical exercise applications.
• Existing methods face challenges with diverse sensor data shapes and achieving high accuracy.
• Convolutional Neural Networks (CNNs) offer potential for sophisticated pattern recognition in sensor data.

Purpose of the Study:

• To develop a novel human activity recognition (HAR) system using CNNs tailored for individual sensor types.
• To enhance HAR accuracy by effectively processing diverse data shapes from accelerometers, gyroscopes, and barometers.
• To compare the performance of the proposed CNN-based approach against a standard Support Vector Machine (SVM) classifier.

Main Methods:

• Individual CNN models were designed for each sensor type (accelerometer, gyroscope, barometer) to capture sensor-specific characteristics.
• A late-fusion technique was utilized to combine predictions from the individual CNN models for comprehensive activity classification.
• The proposed CNN approach was benchmarked against a conventional SVM classifier using the one-vs-rest methodology.

Main Results:

• The late-fusion CNN model achieved significantly higher accuracy, with validation accuracy of 99.35% and final test accuracy of 94.83%.
• The conventional SVM classifier achieved lower accuracies of 87.07% (validation) and 83.10% (final test).
• Combining multiple sensors, a barometer, and a filter algorithm demonstrably improved human movement pattern identification.

Conclusions:

• The proposed CNN-based HAR approach, utilizing sensor-specific models and late fusion, substantially outperforms traditional methods.
• Tailoring CNN architectures to individual sensor data characteristics is key to enhancing HAR accuracy.
• This advanced technique offers a promising solution for accurate human motion detection in various applications.