Multi-Biometric Feature Extraction from Multiple Pose Estimation Algorithms for Cross-View Gait Recognition

Area of Science:

• Biometrics
• Computer Vision
• Pattern Recognition

Background:

• Gait recognition identifies individuals by walking patterns, crucial for long-distance identification.
• Traditional appearance-based methods struggle with real-world variations (clothing, objects, illumination).
• Model-based methods using skeletal key points offer robustness but historically underperform appearance-based methods.

Purpose of the Study:

• To bridge the performance gap between skeleton-based and appearance-based gait recognition.
• To introduce a multi-biometric framework leveraging multiple human pose estimation (HPE) algorithms.
• To enhance the robustness and accuracy of skeleton-based gait recognition.

Main Methods:

• Utilized state-of-the-art HPE algorithms (OpenPose, AlphaPose, HRNet) to generate diverse skeleton data from single videos.
• Employed a residual graph convolutional network (ResGCN) for feature extraction from skeleton data.
• Implemented feature-level fusion (FLF) and decision-level fusion (DLF) techniques.

Main Results:

• The multi-biometric framework demonstrated superior skeleton-based cross-view gait recognition.
• Achieved state-of-the-art performance on the CASIA-B dataset.
• FLF aggregated features point-wise, while DLF used majority voting for final decisions.

Conclusions:

• The proposed multi-biometric framework effectively enhances skeleton-based gait recognition accuracy.
• Leveraging multiple HPE algorithms and fusion techniques overcomes limitations of single-source methods.
• This approach offers a robust solution for real-world gait recognition challenges.