GaitSet: Cross-View Gait Recognition Through Utilizing Gait As a Deep Set.

Area of Science:

• Biometrics
• Computer Vision
• Machine Learning

Background:

• Gait recognition is crucial for security and identification but faces challenges with temporal information and flexibility.
• Existing methods using gait templates or sequences have limitations in preserving gait dynamics and adaptability.

Purpose of the Study:

• To propose a novel deep set perspective for gait recognition.
• To develop a flexible and robust gait identification method immune to frame order and environmental variations.

Main Methods:

• Utilizing a global-local fused deep network to process gait frames as a deep set.
• Developing a method invariant to frame permutations and adaptable to diverse acquisition scenarios.

Main Results:

• Achieved 96.1% rank-1 accuracy on CASIA-B and 87.9% on OU-MVLP under normal conditions.
• Demonstrated high robustness with 90.8% accuracy (bag-carrying) and 70.3% (coat-wearing) on CASIA-B.
• Maintained 85.0% accuracy on CASIA-B with only 7 frames, showing effectiveness with limited data.

Conclusions:

• The deep set approach offers a flexible and robust gait recognition solution.
• The proposed method significantly outperforms existing techniques, especially under varied and challenging conditions.
• This approach holds promise for real-world applications in security and identification.