A dual-branch pedestrian re-identification method CPHMNet based on multi-dimensional feature fusion and integrated...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Pedestrian re-identification (re-ID) faces challenges due to pose variations and occlusions, impacting recognition accuracy.
• Existing methods struggle to effectively extract discriminative features under these adverse conditions.

Purpose of the Study:

• To propose CPHMNet, a novel pedestrian re-identification method.
• To enhance feature discrimination and fusion for improved recognition performance.
• To address interferences caused by pose variations and occlusions in pedestrian re-ID tasks.

Main Methods:

• Developed a novel MDA module to improve feature discrimination for similar pedestrians.
• Employed a dual-branch network: one for pose estimation and local features, another with CBAM attention for information extraction.
• Utilized a multi-dimensional feature fusion network (MDFF) for weighted fusion and a joint loss function for model constraints.

Main Results:

• Achieved mAP of 89.9% on Market1501, 81.1% on DukeMTMC-reID, and 61.6% on MSMT17.
• Reached Rank-1 accuracies of 95.9% on Market1501, 90.3% on DukeMTMC-reID, and 82.5% on MSMT17.
• Demonstrated superior performance in practical, real-world pedestrian re-identification scenarios.

Conclusions:

• CPHMNet effectively overcomes pose variations and occlusions in pedestrian re-identification.
• The proposed multi-dimensional feature fusion and integrated pose estimation significantly boost recognition performance.
• The method shows strong potential for real-world applications requiring robust pedestrian re-identification.