Human Activity Recognition Using Deep Residual Convolutional Network Based on Wearable Sensors.

Area of Science:

• Biomedical and Health Informatics
• Artificial Intelligence
• Machine Learning

Background:

• Human Activity Recognition (HAR) is crucial for monitoring daily activities and health behaviors.
• Accurate HAR provides insights into physical activity for chronic disease management and lifestyle promotion.

Purpose of the Study:

• To propose DKInception, a novel deep learning model for enhanced human activity recognition.
• To evaluate DKInception's performance against existing models on benchmark datasets.

Main Methods:

• Developed DKInception, integrating deep convolutional residual networks with an attention mechanism.
• Employed multi-scale convolution kernels and the Inception ResNet architecture for efficient temporal feature extraction.
• Conducted extensive experiments on four benchmark HAR datasets: UCI-HAR, Opportunity, Daphnet, and PAMAP2.

Main Results:

• DKInception demonstrated superior performance compared to existing models across multiple evaluation metrics.
• Achieved high accuracy rates: 95.70% (UCI-HAR), 87.48% (Opportunity), 94.00% (Daphnet), and 89.72% (PAMAP2).

Conclusions:

• DKInception offers effective, fast convergence, and robust scaling properties for HAR tasks.
• The proposed model significantly advances the capabilities of human activity recognition in health informatics.