Overcoming occlusion in person re-identification: A multi-level attention transformer approach
Person re-identification (ReID) in real-world surveillance scenarios is a very challenging problem where occlusions are a major culprit that can significantly degrade the performance of current systems. In this paper, we take a step closer to solve this important problem by introducing a novel Multi...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | en |
| Published: |
Mehran University of Engineering and Technology
2026
|
| Online Access: | http://eprints.utem.edu.my/id/eprint/29508/2/028421401202611227.pdf http://eprints.utem.edu.my/id/eprint/29508/ https://murjet.muet.edu.pk/index.php/home/article/view/449 https://doi.org/10.22581/0449 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Person re-identification (ReID) in real-world surveillance scenarios is a very challenging problem where occlusions are a major culprit that can significantly degrade the performance of current systems. In this paper, we take a step closer to solve this important problem by introducing a novel Multi-Level Attention Mechanism (MLAM) for occluded person re-identification. The method combines spatial, channel, and global context attention in order to handle various occlusions from partial to severe. The proposed method integrates two significant architectures, the Multi-Level Attention Transformer Network (MLATN) and the Occlusion-Aware ReID Transformer (OART). In particular, we show that the proposed framework can achieve adaptive feature extraction and occlusion-aware fusion, which leads to large robustness improvement when applied for adaptive ReID in real-world challenging environments. This study examines the approach on several large relevant datasets,
Occluded-DukeMTMC and Occluded-REID, and shows that the approach outperforms previous methods. For the Occluded DukeMTMC, the MLAM achieves state-of-the-art performance, achieving 2.7% and 5.1% in Rank-1 accuracy and mean Average Precision (mAP), respectively. At the same time, we introduced a new model-invariant metric named Occlusion Robustness Index (ORI) to quantify model robustness to occlusion. Aside from surveillance, the research findings have implications for autonomous driving, robotics, and augmented reality. Nevertheless, there have been tremendous advances in this area, which illuminate important ethical concerns surrounding privacy and information protection and a need for responsible development and implementation of such technologies. As such, we believe this work represents a significant step towards occluded person re-identification and the
achievement of robust, adaptable visual recognition systems for real-world environments. |
|---|