Adaptive Gait Feature Learning Using Mixed Gait Sequence

Area of Science:

• Computer Vision
• Biometrics
• Pattern Recognition

Background:

• Gait recognition offers contactless identification but struggles with occlusions (coats, backpacks) and viewpoint variations.
• Occluded gait silhouettes and diverse camera angles significantly degrade recognition performance.

Purpose of the Study:

• To enhance gait recognition accuracy under occlusion and varying camera viewpoints.
• To address challenges posed by coats (CL), backpacks (BG), and different camera angles in gait identification.

Main Methods:

• Proposed Gait Sequence Mixing (GSM) for data augmentation to mitigate information loss from occlusions.
• Introduced Multigranularity Feature Extraction (MFE) to capture diverse gait features from incomplete silhouettes.
• Developed Feature Distance Alignment (FDA) to refine features and reduce viewpoint variance.

Main Results:

• The proposed methods, including GSM, MFE, and FDA, significantly improved gait recognition performance.
• Experiments on CASIA-B and mini-OUMVLP datasets showed superior results compared to state-of-the-art methods.
• Integrating GSM and FDA modules boosted the accuracy of existing state-of-the-art gait recognition techniques.

Conclusions:

• The combined approach effectively tackles occlusion and viewpoint challenges in gait recognition.
• The proposed methods offer a robust solution for reliable human identification using gait analysis.
• This work advances the practical applicability of gait recognition in real-world surveillance and security systems.