A human activity recognition model based on deep neural network integrating attention mechanism

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Human Activity Recognition (HAR) is vital across various domains.
• Current HAR methods, including manual feature extraction, codebook-based approaches, and deep learning, possess inherent limitations.
• There is a need for advanced HAR models that are accurate and generalizable.

Purpose of the Study:

• To develop a novel deep neural network, DCAM-Net (DeepConvAttentionMLPNet), for Human Activity Recognition.
• To evaluate DCAM-Net's performance without relying on pre-trained model weights.
• To demonstrate the efficacy of integrating Convolutional Neural Networks (CNNs) and Multilayer Perceptrons (MLPs) with an attention mechanism for HAR.

Main Methods:

• DCAM-Net was developed, integrating CNN and MLP architectures with a dual attention mechanism.
• The model was trained and validated using smartphone sensor data (acceleration and gyroscope) from 30 participants.
• Key architectural components include multi-scale feature extraction, residual and skip connections, dropout, batch normalization, and the AdamW optimizer.

Main Results:

• DCAM-Net achieved a high average accuracy of 99.03% via five-fold cross-validation.
• The proposed model demonstrated superior performance compared to other existing HAR models.
• The study confirmed DCAM-Net's robust generalization ability on unseen data.

Conclusions:

• DCAM-Net represents a significant advancement in deep learning-based Human Activity Recognition.
• The model's architecture effectively leverages attention mechanisms and CNN-MLP integration for high accuracy.
• Future research should focus on diverse datasets, real-time efficiency, and sensor fusion for practical HAR applications.