Monocular 3-D gait tracking in surveillance scenes.

Area of Science:

• Computer Vision
• Biometrics
• Human Pose Estimation

Background:

• Gait recognition offers noninvasive, remote biometric authentication.
• Surveillance scenarios present challenges like multiple individuals, camera distance, and angle.
• Existing systems struggle with view-invariant gait recognition.

Purpose of the Study:

• To develop a view-invariant monocular 3D human pose tracking methodology.
• To enhance gait recognition performance in realistic surveillance settings.
• To enable accurate 3D pose tracking despite significant camera perspective effects.

Main Methods:

• Modeling 3D body poses and camera viewpoints using a low-dimensional manifold.
• Learning a generative silhouette model from the manifold to training views.
• Employing recursive Bayesian sampling for joint tracking over ground plane and pose-viewpoint manifold.
• Utilizing homographic transformation for silhouette projection and matching.

Main Results:

• Successful tracking of 3D human walking poses within a 4D state space.
• Demonstrated significant improvements of homographic alignment over similarity transformation.
• Provided quantitative pose tracking results on the CAVIAR dataset with high perspective effects.

Conclusions:

• The proposed methodology achieves robust 3D human pose tracking in challenging, real-world surveillance conditions.
• View-invariant tracking is crucial for effective gait recognition in unconstrained environments.
• The approach offers a promising solution for advanced biometric authentication systems.