Human Activity Recognition Using Attention-Mechanism-Based Deep Learning Feature Combination.

Area of Science:

• Computer Science
• Artificial Intelligence
• Signal Processing

Background:

• Human Activity Recognition (HAR) is crucial for healthcare, elder care, and monitoring.
• Deep Learning (DL) models, particularly Convolutional Neural Networks (CNNs), excel at automatic feature extraction from sensor data for HAR.
• Existing HAR methods often rely on hand-crafted features, which can be complex and suboptimal.

Purpose of the Study:

• To propose a novel deep learning methodology for enhanced Human Activity Recognition (HAR).
• To improve HAR accuracy by combining multi-stage features and incorporating an attention mechanism.
• To develop a generalized model structure for sensor-based HAR using Convolutional Neural Networks (CNNs) and CBAM modules.

Main Methods:

• Utilized Convolutional Neural Networks (CNNs) for HAR, processing raw sensor signals as spectrograms.
• Implemented a novel approach combining features from multiple convolutional stages for richer representations.
• Integrated a Convolutional Block Attention Module (CBAM) to refine feature extraction and enhance model accuracy.

Main Results:

• The proposed CNN model achieved high classification accuracies: 96.86% on KU-HAR, 93.48% on UCI-HAR, and 93.89% on WISDM.
• The methodology demonstrated superior performance compared to previous HAR approaches across multiple evaluation metrics.
• The integration of multi-stage feature combinations and attention mechanisms proved effective for HAR.

Conclusions:

• The developed deep learning model offers a comprehensive and competent solution for sensor-based Human Activity Recognition.
• The novel feature extraction technique, leveraging multi-stage feature combinations and attention, significantly boosts HAR accuracy.
• This generalized model structure shows promise for diverse HAR applications in healthcare and monitoring.