Gait recognition using spatio-temporal representation fusion learning network with IMU-based skeleton graph and body...

Area of Science:

• Biomedical Engineering
• Computer Science
• Robotics

Background:

• Precise recognition of human lower limb movements using wearable sensors is crucial for human-computer interaction.
• Existing methods often overlook dynamic spatial information, limiting decoding accuracy and robustness.
• Inertial Measurement Unit (IMU) sensors offer a viable solution for capturing movement data.

Purpose of the Study:

• To develop a deep learning model that effectively utilizes spatial and temporal information from IMU-based skeleton data for human lower limb movement recognition.
• To address the limitations of existing methods in decoding accuracy and robustness.
• To establish a benchmark for IMU-based human lower limb movement recognition.

Main Methods:

• Constructed skeleton graph data from IMU sensors.
• Proposed a two-branch deep learning model, TCNN-MGCHN, integrating temporal and graph convolutional modules.
• Developed a multi-scale graph convolutional module with a body partitioning strategy for spatial feature extraction.
• Incorporated an attention sub-module for enhanced temporal feature extraction.

Main Results:

• The TCNN-MGCHN model demonstrated superior classification performance compared to mainstream methods on a self-constructed dataset.
• The model effectively mined meaningful spatial and temporal feature representations from IMU-based skeleton graph data.
• Ablation studies confirmed the effectiveness of the proposed model components.

Conclusions:

• TCNN-MGCHN offers a robust and accurate approach for human lower limb movement recognition using IMU sensors.
• The model's ability to capture spatio-temporal features provides a significant advancement in the field.
• This study provides a valuable benchmark for future research in IMU-based movement recognition and deep learning applications.