GaitRGA: Gait Recognition Based on Relation-Aware Global Attention

Area of Science:

• Biometrics
• Computer Vision
• Artificial Intelligence

Background:

• Gait recognition, a non-invasive biometric technique, faces performance limitations in real-world applications due to viewing angle variations, occlusion, and clothing changes.
• Existing methods often struggle with dynamic scenarios, necessitating advancements beyond traditional local convolutional approaches.

Purpose of the Study:

• To propose a novel gait recognition method that addresses the challenges of real-world applications.
• To enhance the accuracy and robustness of gait recognition systems by capturing global structural information.

Main Methods:

• Introduction of a Relational-aware Global Attention (RGA) module for gait recognition.
• Utilizing a shallow convolutional model to stack pairwise feature associations and learn attention, capturing global structural information within gait sequences.
• Overcoming limitations of solely relying on local convolutions by incorporating global relationships.

Main Results:

• The proposed GaitRGA method demonstrates significant performance improvements on multiple datasets (Grew, Gait3D, SUSTech1k).
• The RGA module effectively captures global structural information, aiding in more precise attention learning and improved gait feature differentiation.
• Enhanced recognition performance, particularly in complex and dynamic real-world scenarios.

Conclusions:

• The relational-aware global attention approach offers a promising solution for improving gait recognition in challenging, real-world environments.
• GaitRGA effectively handles variations in viewing angle, occlusion, and clothing, leading to more reliable biometric identification.
• This method enhances the differentiation of gait features by leveraging global structural information inherent in human walking.