Heterogeneous feature-aware Transformer-CNN coupling network for person re-identification.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Smart City Technology

Background:

• Person re-identification is crucial for smart city infrastructure.
• Current methods struggle with local/global feature integration and cluttered backgrounds.

Purpose of the Study:

• To propose a novel Transformer-CNN Coupling Network (TCCNet) for improved person re-identification.
• To effectively capture fluctuant body region features and handle heterogeneous data.

Main Methods:

• Developed TCCNet, a hybrid network combining Transformer and CNN architectures.
• Introduced Low-level Feature Coupling Module (LFCM) and High-level Feature Coupling Module (HFCM) for feature complementarity.
• Utilized duplicate loss to incorporate semantic information from both network branches.

Main Results:

• TCCNet demonstrated superior performance on large-scale person re-identification benchmarks.
• Achieved competitive results against state-of-the-art approaches.
• On the MSMT17 dataset, TCCNet reached 66.9% mAP and 84.5% Rank-1 accuracy.

Conclusions:

• TCCNet effectively integrates local and global features for robust person re-identification.
• The proposed network reduces the impact of cluttered backgrounds on identification accuracy.
• TCCNet offers a promising solution for smart city surveillance systems.