Human Activity Recognition via Attention-Augmented TCN-BiGRU Fusion

Area of Science:

• Wearable sensor technology
• Deep learning for human activity recognition (HAR)

Background:

• Deep learning-based HAR faces challenges in multi-scale feature extraction and noise robustness.
• Effective feature extraction from multi-source wearable sensor data is crucial for accurate human activity recognition.

Purpose of the Study:

• To propose the TGA-HAR model for enhanced human activity recognition.
• To address challenges in multi-scale temporal feature extraction and robustness in noisy, multi-source data.

Main Methods:

• Developed the TGA-HAR model integrating Temporal Convolutional Neural Networks (TCN) and Bidirectional Gated Recurrent Units (BiGRU).
• Employed TCN with dilated convolutions for multi-order temporal feature extraction.
• Utilized BiGRU for bidirectional temporal contextual correlation and residual connections for gradient stability.
• Incorporated an adaptive weighted attention mechanism to enhance feature representation.

Main Results:

• Achieved high test accuracies: 99.37% on WISDM, 95.36% on USC-HAD, and 96.96% on PAMAP2 datasets.
• Demonstrated robustness on real-world collected datasets.
• The TGA-HAR model shows significant improvements in complex human activity recognition.

Conclusions:

• The TGA-HAR model offers a robust and effective solution for complex human activity recognition using wearable sensors.
• The hierarchical feature abstraction and attention mechanism contribute to improved performance and robustness.