Human Activity Recognition Based on Residual Network and BiLSTM.

Area of Science:

• Computer Science
• Biomedical Engineering
• Machine Learning

Background:

• Human Activity Recognition (HAR) is crucial for applications in sports and health.
• Existing deep learning models often struggle with effective spatial and temporal feature extraction from human activity data.

Purpose of the Study:

• To propose a novel deep learning model for enhanced human activity recognition.
• To address the limitations of existing models in extracting spatial and temporal features.

Main Methods:

• A deep learning model integrating residual blocks for spatial feature extraction and bi-directional Long Short-Term Memory (BiLSTM) for temporal dependencies.
• Utilized MEMS inertial sensor data for multidimensional signal processing.
• Developed a custom dataset of six common human activities and validated on public datasets (WISDM, PAMAP2).

Main Results:

• The proposed model achieved high accuracy: 96.95% on the custom dataset, 97.32% on WISDM, and 97.15% on PAMAP2.
• Demonstrated superior performance and a reduced number of parameters compared to existing HAR models.
• Successfully extracted spatial and temporal features from human activity data.

Conclusions:

• The developed deep learning model effectively recognizes human activities by integrating residual blocks and BiLSTM.
• The model offers improved performance and efficiency over existing methods for HAR.
• This approach holds significant potential for applications in sports and health monitoring.