Scholarlink

<p>Traffic object detection is crucial for autonomous driving. Current object detection algorithms face challenges in detecting occluded objects and maintaining accuracy under varying lighting conditions. This paper proposes an improved single-stage object detector based on the YOLOv7 method, incorporating Progressive Feature Fusion (PFF) and Reverse Attention (RA) networks. The PFF module progressively refines and fuses feature maps across scales and layers, enhancing the feature representation for the neck network and strengthening inter-layer connectivity. The RA module reverses the output of deep feature maps to guide the network in discovering and supplementing details, improving detection accuracy. Extensive experiments on the KITTI and PASCAL VOC datasets demonstrate the effectiveness of our approach. Our method achieves 73.8% and 62.9% <i>mAP</i>@0.5:0.95 on KITTI and PASCAL VOC, respectively, outperforming the baseline YOLOv7 by 4.3% and 1.9%. This paper shows that the proposed PFF and RA mechanisms significantly improve detection accuracy, particularly for occluded and small objects in traffic scenes.</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

A method of object detection network with progressive feature fusion and reverse attention for traffic scenes

  • 期刊论文

  • 发表日期：

    2025年7月30日

查看全文
  • Zhong Qu,
  • Haoming Qu,
  • Haonan Yin,
  • Xuejuan Han,
  • Shufang Xia

摘要

Traffic object detection is crucial for autonomous driving. Current object detection algorithms face challenges in detecting occluded objects and maintaining accuracy under varying lighting conditions. This paper proposes an improved single-stage object detector based on the YOLOv7 method, incorporating Progressive Feature Fusion (PFF) and Reverse Attention (RA) networks. The PFF module progressively refines and fuses feature maps across scales and layers, enhancing the feature representation for the neck network and strengthening inter-layer connectivity. The RA module reverses the output of deep feature maps to guide the network in discovering and supplementing details, improving detection accuracy. Extensive experiments on the KITTI and PASCAL VOC datasets demonstrate the effectiveness of our approach. Our method achieves 73.8% and 62.9% mAP@0.5:0.95 on KITTI and PASCAL VOC, respectively, outperforming the baseline YOLOv7 by 4.3% and 1.9%. This paper shows that the proposed PFF and RA mechanisms significantly improve detection accuracy, particularly for occluded and small objects in traffic scenes.