Scholarlink

<p>Multi-modal vehicle re-identification (ReID) is a critical enabling technology for realizing 24-hour all-weather intelligent transportation surveillance, addressing the severe performance degradation of traditional single-modality RGB methods under varying illumination and adverse weather conditions. However, existing multi-modal ReID methods face three fundamental challenges: blurred structural edges in low-quality images, insufficient local discriminative feature extraction when transferring large-scale pre-trained models to vehicle ReID tasks, and the lack of an explicit suppression mechanism for low-quality modalities in static fusion frameworks. To tackle these challenges, this paper proposes a Quality-Prior Mamba Fusion Network (QPMFNet) for robust multi-modal vehicle ReID across RGB, near-infrared (NIR), and thermal infrared (TIR) modalities. At the input stage, a Modal-Aware Laplacian Edge Enhancement Module (MALEE) is designed to adaptively adjust Laplacian kernel parameters according to the unique imaging characteristics of each modality, achieving effective edge enhancement without additional annotations. Taking the image encoder of Contrastive Language-Image Pre-training (CLIP) as the shared backbone, a Hierarchical Local-Progressive Adapter Model (HLPA) is inserted in parallel to balance pre-trained global semantic knowledge and task-specific local discriminative details through layer-wise progressive token adjustment. Most importantly, a Quality-Prior-guided Mamba Fusion Module (QP-Mamba) is proposed, which for the first time integrates explicit modality quality assessment into the Mamba state space modeling process, enabling selective multi-modal feature fusion and efficient long-range dependency modeling. Extensive experiments on three public benchmark datasets (WMVeID863, RGBNT100, and MSVR310) demonstrate that the proposed method achieves mAP scores of 70.6%, 85.6%, and 50.1% respectively, outperforming the current state-of-the-art methods by 0.8%, 1.2%, and 2.3% mAP on the corresponding datasets. The results validate the superior effectiveness and generalization capability of QPMFNet in complex real-world surveillance scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Quality-prior mamba fusion network for multi-modal vehicle re-identification

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Xiaolu Cui,
  • Rui Zhang,
  • Jiabao Wang,
  • Yang Li,
  • Zhuang Miao

摘要

Multi-modal vehicle re-identification (ReID) is a critical enabling technology for realizing 24-hour all-weather intelligent transportation surveillance, addressing the severe performance degradation of traditional single-modality RGB methods under varying illumination and adverse weather conditions. However, existing multi-modal ReID methods face three fundamental challenges: blurred structural edges in low-quality images, insufficient local discriminative feature extraction when transferring large-scale pre-trained models to vehicle ReID tasks, and the lack of an explicit suppression mechanism for low-quality modalities in static fusion frameworks. To tackle these challenges, this paper proposes a Quality-Prior Mamba Fusion Network (QPMFNet) for robust multi-modal vehicle ReID across RGB, near-infrared (NIR), and thermal infrared (TIR) modalities. At the input stage, a Modal-Aware Laplacian Edge Enhancement Module (MALEE) is designed to adaptively adjust Laplacian kernel parameters according to the unique imaging characteristics of each modality, achieving effective edge enhancement without additional annotations. Taking the image encoder of Contrastive Language-Image Pre-training (CLIP) as the shared backbone, a Hierarchical Local-Progressive Adapter Model (HLPA) is inserted in parallel to balance pre-trained global semantic knowledge and task-specific local discriminative details through layer-wise progressive token adjustment. Most importantly, a Quality-Prior-guided Mamba Fusion Module (QP-Mamba) is proposed, which for the first time integrates explicit modality quality assessment into the Mamba state space modeling process, enabling selective multi-modal feature fusion and efficient long-range dependency modeling. Extensive experiments on three public benchmark datasets (WMVeID863, RGBNT100, and MSVR310) demonstrate that the proposed method achieves mAP scores of 70.6%, 85.6%, and 50.1% respectively, outperforming the current state-of-the-art methods by 0.8%, 1.2%, and 2.3% mAP on the corresponding datasets. The results validate the superior effectiveness and generalization capability of QPMFNet in complex real-world surveillance scenarios.