TCN-attention-HAR: human activity recognition based on attention mechanism time convolutional network

Area of Science:

• * Computer Science
• * Biomedical Engineering
• * Human-Computer Interaction

Background:

• * Wearable sensors are crucial for medical applications and human-computer interaction due to portability and privacy.
• * Human activity recognition from sensor data is vital for these fields, necessitating improved performance.
• * Existing models face challenges in time-varying feature extraction and gradient issues in deep networks.

Purpose of the Study:

• * To propose an improved human activity recognition model addressing limitations in temporal feature extraction and network depth.
• * To enhance the recognition performance of human activities using wearable sensor data.
• * To validate the model's effectiveness on public datasets and through knowledge distillation.

Main Methods:

• * Development of a Temporal Convolutional Network (TCN) integrated with an attention mechanism (TCN-Attention-HAR).
• * Optimization of TCN's temporal feature extraction via appropriate receiver domain sizing.
• * Application of attention mechanisms to prioritize critical feature information for improved learning.

Main Results:

• * Achieved performance improvements of 1.13% (WISDM), 1.83% (PAMAP2), and 0.51% (USC-HAD) over existing advanced models.
• * Demonstrated excellent recognition performance on multiple open datasets.
• * In knowledge distillation, the student model (0.1% of teacher parameters) showed improved accuracy, even surpassing the teacher model by 0.14% on WISDM.

Conclusions:

• * The TCN-Attention-HAR model effectively extracts temporal features and emphasizes key information for superior human activity recognition.
• * The proposed model offers a significant advancement in wearable sensor-based activity recognition.
• * Knowledge distillation experiments highlight the model's efficiency and potential for deployment in resource-constrained environments.