CLASH: Complementary Learning With Neural Architecture Search for Gait Recognition.

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Gait recognition identifies individuals by walking patterns, often using silhouette analysis.
• Silhouette-based methods are limited by sparse boundary representations, lacking dense spatial-temporal information.
• Existing methods struggle with sensitivity to subtle walking variations and robustness in real-world conditions.

Purpose of the Study:

• To enhance gait recognition by improving sensitivity to walking patterns while maintaining robustness.
• To introduce a novel framework that combines a dense texture descriptor with a neural architecture search for complementary learning.

Main Methods:

• Developed the Complementary Learning with neural Architecture SearcH (CLASH) framework.
• Introduced the dense spatial-temporal field (DSTF) to represent walking patterns using dense, pixel-level texture.
• Implemented neural architecture search based complementary learning (NCL) to integrate DSTF with silhouette information.

Main Results:

• Achieved high rank-1 accuracy on benchmark datasets: 98.8% (CASIA-B, condition 1), 96.5% (CASIA-B, condition 2), 89.3% (CASIA-B, condition 3), and 91.9% (OU-MVLP).
• Significantly outperformed state-of-the-art silhouette-based methods on in-the-wild datasets (Gait3D by 16.3%, GREW by 19.7%).
• Demonstrated effectiveness in both controlled (in-the-lab) and uncontrolled (in-the-wild) environments.

Conclusions:

• The proposed CLASH framework, leveraging DSTF and NCL, effectively enhances gait recognition accuracy and robustness.
• This approach overcomes limitations of sparse silhouette representations by incorporating dense texture information.
• CLASH represents a significant advancement for identifying individuals based on their unique walking patterns.