Cloth-Changing Person Re-Identification Based on the Backtracking Mechanism
Xuan Liu, Hua Han, Kaiyu Xu, Li Qian Huang · IEEE Access · 2025
Person re-identification has been extensively studied and has made significant progress in recent years. However, traditional person re-identification methods mainly rely on clothing-related features, which become unreliable in real-world scenarios involving frequent clothing changes. To address this issue, we propose a novel cloth-changing person re-identification method based on a backtracking mechanism, which extracts identity features independent of clothing and mitigates the impact of clothing changes on recognition performance. Unlike traditional methods, our approach simulates the brain’s cognitive mechanism, effectively combining appearance features and body shape information to achieve a deeper understanding of identity. Additionally, we introduce a feature infiltration module that integrates body shape features with appearance features extracted by the backbone network, enabling the model to extract clothing-independent features. During training, we use a clothes suppression loss function to reduce the model’s reliance on clothing features, further improving robustness. In the inference stage, the backtracking branch is removed, allowing the main network to focus on extracting clothing-independent features, thereby improving recognition accuracy. Experiments on two benchmark datasets demonstrate that our approach has better robustness and accuracy, thereby confirming its superiority.