GaitTriViT and GaitVViT: Transformer-based methods emphasizing spatial or temporal aspects in gait recognition.

Area of Science:

• Computer Vision
• Biometrics
• Artificial Intelligence

Background:

• Gait recognition is a promising biometric technology, but challenges persist with low-resolution and long-distance subject identification.
• Traditional methods often rely on Convolutional Neural Networks (CNNs), which may not fully capture complex spatio-temporal gait features.
• Vision Transformers (ViTs) have emerged as powerful alternatives in computer vision, offering potential for improved performance.

Purpose of the Study:

• To introduce and evaluate two novel Transformer-based methods for gait recognition: GaitTriViT and GaitVViT.
• To assess the effectiveness of ViTs in extracting fine-grained spatial and temporal features for gait analysis.
• To compare the performance of these new methods against current state-of-the-art (SOTA) approaches.

Main Methods:

• Developed GaitTriViT, a method utilizing Vision Transformers (ViTs) for enhanced spatial feature extraction.
• Developed GaitVViT, a method employing Video Vision Transformers (Video ViT) to improve temporal feature understanding.
• Evaluated the performance of both methods on gait recognition tasks, comparing them with existing SOTA models.

Main Results:

• The proposed Transformer-based methods, GaitTriViT and GaitVViT, demonstrate encouraging performance in gait recognition.
• Results indicate existing gaps and challenges in applying ViTs to gait recognition, particularly with complex scenarios.
• The study highlights both the potential and the difficulties encountered when using ViTs for this biometric task.

Conclusions:

• Transformer-based methods, specifically ViTs, hold significant promise for advancing gait recognition technology.
• Further research and development are needed to overcome the identified challenges and fully leverage the capabilities of ViTs in this domain.
• The future of Vision Transformers in gait recognition appears promising, despite current limitations.