Scholarlink

<p>Accurate identification of various types of objects in road scenes is the core of planning safe and reasonable driving strategies for self-driving vehicles. Complex traffic object detection faces several challenges, including limited datasets, significant visual noise, object occlusion, and the frequent misdetection of small objects. To address these challenges, this study constructs a Cityscapes object detection dataset with a scale of 10000 images, compensating for the lack of object detection data in the existing public dataset. Additionally, an efficient frequency domain learning you only look once detector (FDL-YOLO) is proposed, specifically designed for detecting small objects. The detector incorporates an efficient multi-scale feature extraction and noise filtering module (EMNF), along with a small object differential edge enhancement detection path (DEEP). EMNF, based on partial convolution, leverages shortest and longest gradient path strategies and Fourier transform to enhance feature extraction in low visibility and dense occlusion while optimizing memory access and maximizing computational efficiency. DEEP employs the traditional differential image enhancement methods to improve the model’s sensitivity to perceive small object edge features. To verify the effectiveness and robustness of FDL-YOLO, experiments are conducted using the BDD100k and self-built Cityscapes object detection dataset. Experimental results show that FDL-YOLO, with an FPS of 51, meets real-time performance requirements. On two datasets, the mAP@50 detection metric improved by 7.7% and 6.0%, respectively. Comparative experiments verify our model’s performance superiority over the latest and existing popular object detection models. The dataset can be accessed via the following link: <a href="https://zenodo.org/records/14847722">https://zenodo.org/records/14847722</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

FDL-YOLO: efficient real-time traffic detection network based on frequency domain learning

  • 期刊论文

  • 发表日期：

    2025年6月5日

查看全文
  • Hanhuang Ling,
  • Ling Li,
  • Duotao Pan,
  • Yuxiang Zhang

摘要

Accurate identification of various types of objects in road scenes is the core of planning safe and reasonable driving strategies for self-driving vehicles. Complex traffic object detection faces several challenges, including limited datasets, significant visual noise, object occlusion, and the frequent misdetection of small objects. To address these challenges, this study constructs a Cityscapes object detection dataset with a scale of 10000 images, compensating for the lack of object detection data in the existing public dataset. Additionally, an efficient frequency domain learning you only look once detector (FDL-YOLO) is proposed, specifically designed for detecting small objects. The detector incorporates an efficient multi-scale feature extraction and noise filtering module (EMNF), along with a small object differential edge enhancement detection path (DEEP). EMNF, based on partial convolution, leverages shortest and longest gradient path strategies and Fourier transform to enhance feature extraction in low visibility and dense occlusion while optimizing memory access and maximizing computational efficiency. DEEP employs the traditional differential image enhancement methods to improve the model’s sensitivity to perceive small object edge features. To verify the effectiveness and robustness of FDL-YOLO, experiments are conducted using the BDD100k and self-built Cityscapes object detection dataset. Experimental results show that FDL-YOLO, with an FPS of 51, meets real-time performance requirements. On two datasets, the mAP@50 detection metric improved by 7.7% and 6.0%, respectively. Comparative experiments verify our model’s performance superiority over the latest and existing popular object detection models. The dataset can be accessed via the following link: https://zenodo.org/records/14847722.