Gait Recognition with Self-Supervised Learning of Gait Features Based on Vision Transformers.

Area of Science:

• Computer Science
• Biometrics
• Machine Learning

Background:

• Gait recognition is a valuable biometric for surveillance and access control due to its unique, non-cooperative, and difficult-to-obscure nature.
• Current gait recognition methods predominantly use supervised learning and Convolutional Neural Networks (CNNs), facing challenges with data annotation costs and the under-exploration of newer architectures like Vision Transformers (ViT).

Purpose of the Study:

• To propose a novel self-supervised learning (SSL) approach for pretraining gait feature extractors using the Vision Transformer (ViT) architecture.
• To address the limitations of supervised learning in gait recognition, specifically the expense and time associated with data annotation.
• To explore the potential of ViT in gait recognition, an area less researched compared to CNNs.

Main Methods:

• Implemented a self-supervised learning (SSL) strategy using the DINO model for pretraining.
• Utilized the Vision Transformer (ViT) architecture for automatic feature extraction from gait data.
• Employed a supervised approach to train a fully connected neural network classifier on the pre-extracted gait features.

Main Results:

• Demonstrated the effectiveness of the proposed SSL approach for gait feature extraction.
• Achieved successful gait recognition by pretraining ViT feature extractors without extensive manual annotation.
• Validated the approach on benchmark datasets, including CASIA-B and OU-MVLP.

Conclusions:

• The proposed self-supervised learning method effectively learns gait features using Vision Transformers.
• This approach offers a viable alternative to traditional supervised methods, reducing annotation dependency.
• The study highlights the potential of ViT in advancing gait recognition technology.