Inertial-Measurement-Unit-Based Novel Human Activity Recognition Algorithm Using Conformer.

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Inertial-measurement-unit (IMU)-based human activity recognition (HAR) performance has advanced with new classification models.
• Transformer models offer efficient temporal dependency extraction but can be enhanced with convolutional neural network (CNN) layers for local feature extraction.

Purpose of the Study:

• To introduce and evaluate the conformer model, a novel approach for improving transformer-based HAR models.
• To enhance HAR performance by combining the strengths of transformer models and CNNs.

Main Methods:

• The conformer model, adapted from speech recognition, was applied to HAR tasks.
• Data augmentation using synthetic minority oversampling technique (SMOTE) was employed.
• Performance was evaluated on WISDM, UCI-HAR, and PAMAP2 datasets using baseline transformer and 1D-CNN models for comparison.

Main Results:

• The conformer-based HAR model demonstrated superior performance compared to baseline transformer-based and 1D-CNN models.
• The proposed algorithm outperformed recent studies that did not incorporate recurrent neural network (RNN) series components.

Conclusions:

• The conformer model represents a significant advancement in IMU-based HAR, outperforming existing methods.
• Integrating CNN layers into transformer architectures effectively improves the recognition of human activities.