Coupled Patch Alignment for Matching Cross-view Gaits

Area of Science:

• Computer Vision
• Biometrics
• Pattern Recognition

Background:

• Gait recognition is valuable for identification, even at a distance and low resolution.
• View changes significantly degrade the performance of existing gait recognition systems.
• Robust methods are needed to handle variations in viewpoint for accurate gait analysis.

Purpose of the Study:

• To propose a novel algorithm, Coupled Patch Alignment (CPA), for effective cross-view gait recognition.
• To develop a method that addresses the challenge of view invariance in human gait identification.
• To enhance the discriminative power of gait features across different viewing angles.

Main Methods:

• The Coupled Patch Alignment (CPA) algorithm constructs patches using samples and their nearest neighbors.
• An objective function is designed to balance cross-view intra-class compactness and inter-class separability within each patch.
• Local patches are unified into a global objective function, with an extension to Multi-dimensional Patch Alignment (MPA) for multiple views.

Main Results:

• The proposed CPA algorithm demonstrates effective matching of gait pairs across different views.
• Theoretical analysis reveals a connection between CPA and Canonical Correlation Analysis (CCA).
• Experiments on CASIA(B), USF, and OU-ISIR datasets show superior performance compared to existing methods.

Conclusions:

• The Coupled Patch Alignment (CPA) algorithm significantly improves cross-view gait recognition accuracy.
• The Multi-dimensional Patch Alignment (MPA) extension effectively handles multiple views.
• The proposed methods offer a robust solution for gait recognition challenges posed by viewpoint variations.