Cross-View Person Identification Based on Confidence-Weighted Human Pose Matching

Area of Science:

• Computer Vision
• Biometrics
• Machine Learning

Background:

• Cross-view person identification (CVPI) is challenging due to varying camera angles and potential inaccuracies in 3D pose estimation.
• Current methods primarily rely on appearance and motion features, with pose matching being less effective due to unreliable pose data.

Purpose of the Study:

• To develop a more robust method for CVPI by incorporating accurate 3D human pose information.
• To introduce a novel confidence metric for 3D human pose estimation to improve its utility in CVPI.

Main Methods:

• A new confidence metric was developed for the location of human-body joints in 3D pose estimation.
• A mapping function was proposed to combine pose data with confidence scores, weighting high-confidence joints more heavily.
• Pose information, refined by the confidence metric, was integrated with appearance and motion features.

Main Results:

• The proposed confidence metric demonstrated effectiveness in improving CVPI.
• Integrating pose, appearance, and motion features led to a new state-of-the-art performance in CVPI.
• Experiments on three wearable-camera datasets validated the method's superiority over existing approaches.

Conclusions:

• The novel confidence metric significantly enhances the reliability of 3D pose estimation for CVPI.
• Combining pose, appearance, and motion features through this enhanced pose information yields superior CVPI performance.
• This approach offers a promising direction for improving person identification in complex, real-world scenarios.