Robust Depth-Based Person Re-Identification.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Person re-identification (re-id) commonly relies on RGB appearance, which is susceptible to variations in illumination and clothing.
• Existing RGB-based re-id methods often fail under extreme lighting conditions or when individuals change attire.

Purpose of the Study:

• To enhance person re-identification robustness by incorporating depth information.
• To develop methods that are invariant to illumination and clothing changes.

Main Methods:

• Exploitation of depth voxel covariance descriptor for body shape description.
• Proposal of a locally rotation invariant Eigen-depth feature descriptor.
• Development of a kernelized implicit feature transfer scheme to estimate depth features from RGB images.
• Validation using publicly available depth pedestrian datasets.

Main Results:

• Depth information provides invariant body shape and skeleton data, overcoming RGB limitations.
• The proposed Eigen-depth feature descriptor is locally rotation invariant.
• Combining estimated depth features with RGB appearance features reduces visual ambiguities.
• The proposed methods demonstrate effectiveness compared to existing re-id approaches.

Conclusions:

• Depth information offers a more robust alternative to RGB appearance for person re-identification.
• The Eigen-depth feature and implicit feature transfer scheme significantly improve re-id performance in challenging scenarios.
• This approach enhances the reliability of person re-identification systems in real-world applications.