Enhancing the Transformer Model with a Convolutional Feature Extractor Block and Vector-Based Relative Position...

Area of Science:

• Computer Science
• Artificial Intelligence
• Signal Processing

Background:

• Transformer models are popular for Human Activity Recognition (HAR) due to their ability to capture long-range dependencies.
• Existing Transformer models struggle to effectively utilize complex temporal correlations in Inertial Measurement Unit (IMU) sensor data.

Purpose of the Study:

• To enhance Transformer models for HAR by better utilizing temporal correlations in IMU sensor data.
• To introduce novel components that improve the model's ability to capture both local and global time-series features and relative positional information.

Main Methods:

• Proposed a Convolutional Feature Extractor Block (CFEB) using multi-layer convolutional layers to capture local and global time-series features.
• Introduced Vector-based Relative Position Embedding (vRPE) to better represent temporal positional correlations compared to Absolute Position Embedding (APE).
• Evaluated the enhanced Transformer model on three benchmark HAR datasets: KU-HAR, UniMiB SHAR, and USC-HAD.

Main Results:

• The proposed CFEB and vRPE significantly improved the performance of the Transformer model in HAR tasks.
• The enhanced model demonstrated superior ability in leveraging complex temporal correlations within IMU sensor time-series signals.
• Experimental results confirmed substantial performance gains across multiple HAR datasets.

Conclusions:

• The integration of CFEB and vRPE provides an effective enhancement scheme for Transformer-based HAR.
• The proposed methods enable Transformer models to better utilize a priori information from IMU sensor data for improved activity classification.
• This work offers a promising direction for advancing HAR systems using deep learning models.