Scholarlink

<p>In intelligent surveillance and post-disaster rescue operations, accurate pedestrian detection under low-light conditions remains a critical challenge due to frequent false positives and missed detections. Most existing methods rely on single-stage or static modality fusion schemes, which limit their ability to achieve dynamic cross-modal adaptation and deep semantic complementarity. To address this, we propose a novel two-stage cross-modal fusion framework based on YOLOv11, integrating a Modality-Guided Depth Residual Fusion (MDRF) module at the input stage and a Residual Cross-Modality Fusion (RCMF) module in the mid-backbone. The MDRF module enhances early feature extraction through dual attention mechanisms in channel and spatial dimensions, while the RCMF module achieves semantic alignment and complementary fusion via symmetric cross-attention. Experiments on the ONERA.ROOM dataset demonstrate significant improvements, with Precision increasing by 25.4%, Recall by 16.8%, and mAP@0.5 by 16.3%, achieving an F1 score of 0.48. Here, we show that our approach robustly enhances detection accuracy and stability in complex environments, making it suitable for real-time applications in night-time surveillance and post-disaster response. The proposed dual-stage, bidirectional modality-guided fusion framework provides a robust and adaptive solution for multimodal perception under challenging illumination conditions. The implementation and dataset are publicly available at <a href="https://doi.org/10.5281/zenodo.16809660">https://doi.org/10.5281/zenodo.16809660</a>.</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

Enhancing low-light RGB-D pedestrian detection through dual-stage modality-guided fusion

  • 期刊论文

  • 发表日期：

    2025年12月20日

查看全文
  • Jianhao Guo,
  • Guobing Sun

摘要

In intelligent surveillance and post-disaster rescue operations, accurate pedestrian detection under low-light conditions remains a critical challenge due to frequent false positives and missed detections. Most existing methods rely on single-stage or static modality fusion schemes, which limit their ability to achieve dynamic cross-modal adaptation and deep semantic complementarity. To address this, we propose a novel two-stage cross-modal fusion framework based on YOLOv11, integrating a Modality-Guided Depth Residual Fusion (MDRF) module at the input stage and a Residual Cross-Modality Fusion (RCMF) module in the mid-backbone. The MDRF module enhances early feature extraction through dual attention mechanisms in channel and spatial dimensions, while the RCMF module achieves semantic alignment and complementary fusion via symmetric cross-attention. Experiments on the ONERA.ROOM dataset demonstrate significant improvements, with Precision increasing by 25.4%, Recall by 16.8%, and mAP@0.5 by 16.3%, achieving an F1 score of 0.48. Here, we show that our approach robustly enhances detection accuracy and stability in complex environments, making it suitable for real-time applications in night-time surveillance and post-disaster response. The proposed dual-stage, bidirectional modality-guided fusion framework provides a robust and adaptive solution for multimodal perception under challenging illumination conditions. The implementation and dataset are publicly available at https://doi.org/10.5281/zenodo.16809660.