Feature fusion using deep learning for smartphone based human activity recognition.

Area of Science:

• Computer Science
• Biomedical Engineering
• Machine Learning

Background:

• Human Activity Recognition (HAR) is crucial for personalized health and fitness monitoring.
• HAR model performance relies heavily on extracted features, traditionally requiring domain expertise.
• Deep learning (DL) offers automatic feature extraction, but handcrafted features remain vital.

Purpose of the Study:

• To improve Human Activity Recognition (HAR) model performance.
• To explore the synergistic benefits of combining handcrafted and automatically extracted features.
• To develop a more accurate and robust HAR system.

Main Methods:

• Utilized a hybrid approach combining handcrafted features with DL-based automatic feature extraction.
• Implemented a feature fusion strategy for HAR model input.
• Conducted extensive experiments on both self-collected and public datasets.

Main Results:

• The proposed feature fusion HAR model achieved higher accuracy than state-of-the-art methods.
• Demonstrated superior performance on diverse datasets, indicating generalizability.
• Validated the effectiveness of combining domain knowledge with deep learning approaches.

Conclusions:

• Feature fusion significantly enhances HAR model accuracy.
• The hybrid approach offers a robust solution for human physical activity identification.
• This method holds promise for advancing personalized health and fitness monitoring applications.