View-independent gait events detection using CNN-transformer hybrid network

Area of Science:

• Biomedical Engineering
• Computer Vision
• Machine Learning

Background:

• Accurate gait detection is vital for extracting health information.
• Vision-based gait analysis offers an alternative to sensor-based methods but faces challenges in feature engineering and view dependency.
• Existing methods often require complex feature extraction and are limited to specific viewpoints.

Purpose of the Study:

• To develop a simple, view-independent, and accurate vision-based approach for gait detection.
• To overcome the limitations of traditional feature engineering and lateral view reliance in gait analysis.

Main Methods:

• Deep learning networks, including a 2D convolutional neural network and an attention-based network, were trained using videos of 22 participants performing six distinct gaits.
• The performance of four networks was evaluated on both standard and peculiar gaits using metrics like F1-scores.
• The CNN-Transformer network was specifically investigated for its efficacy.

Main Results:

• All trained networks demonstrated significant detection performance for various gaits.
• The CNN-Transformer network exhibited superior performance for both standard and peculiar gaits.
• The model showed minimal sensitivity to variations in action speed and viewing angles.

Conclusions:

• The CNN-Transformer network presents a highly accurate and view-independent solution for gait detection.
• This vision-based approach simplifies gait analysis, paving the way for broader applications in remote and clinical health monitoring.
• The findings support the integration of advanced deep learning for accessible gait-based health assessments.