MuPeG-The Multiple Person Gait Framework.

Area of Science:

• Computer Vision
• Biometrics
• Pattern Recognition

Background:

• Gait recognition offers uncooperative subject identification.
• Current datasets yield high accuracy (>90%) but lack real-world complexity.
• Existing datasets typically feature only one subject, limiting applicability.

Purpose of the Study:

• To address the limitations of single-subject datasets in gait recognition.
• To introduce a framework (MuPeG) for generating multi-subject augmented datasets.
• To propose an experimental methodology for evaluating gait recognition in complex, multi-subject scenarios.

Main Methods:

• Developed the MuPeG framework to automatically generate multi-subject datasets from existing ones.
• Utilized MuPeG to create augmented versions of TUM-GAID and CASIA-B datasets.
• Conducted experiments on both original (single-subject) and augmented (multi-subject) datasets.

Main Results:

• Accuracy dropped significantly on augmented datasets (54.8% for TUM-GAID, 42.3% for CASIA-B) compared to original datasets (99.7% and 98.0%).
• Demonstrated the substantial increase in difficulty for gait recognition with multiple subjects present.
• Highlighted the inadequacy of current models trained on single-subject data for real-world scenarios.

Conclusions:

• The MuPeG framework effectively generates challenging, realistic multi-subject datasets for gait recognition research.
• Significant performance degradation underscores the need for new models and methodologies.
• The findings pave the way for more robust gait recognition systems applicable in crowded environments.