Scholarlink

<p>The demands of increasing traffic congestion and urbanization have made the development of an intelligent transportation system (ITS) based on smart perception an inevitable trend. In recent years, based on the close integration of computer vision and UAV platforms, the intelligent detection technology of traffic targets from the perspective of UAVs has created an efficient and flexible perception network for the intelligent transportation system. In order to solve problems such as limited resolution, lack of texture, and complicated interference that arise while recognizing small traffic targets from the high-altitude perspective of drones, this study presented an effective end-to-end detector OF-DETR based on the RT-DETR algorithm. First, to better handle complicated background interference in high-altitude detection tasks, we created a Multiscale Global Feature Extraction Module (MGFEM) that enhanced the model’s capacity to model global contextual information. Second, we proposed the Cross-Scale Fusion Encoder, which reduced feature conflicts induced by cross-layer feature fusion and enhanced the representational capacity of detailed information, allowing the decoder to absorb richer feature inputs. In the loss function design, the FMPD-IoU loss function is proposed to direct the model to pay more attention to difficult samples and optimize bounding box regression, while MPDIoU-Aware Query Selection is proposed to offer more precise queries for the decoder. Our approach establishes a strong day-night visual perception foundation for air-ground collaborative intelligent transportation systems, outperforming RT-DETR by 3.36% during the day and 5.19% during the night on mAP50.</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

OF-DETR: an efficient end-to-end detector for tiny traffic targets in aerial images

  • 期刊论文

  • 发表日期：

    2025年12月4日

查看全文
  • Jie Hu,
  • Hanzhang Huang,
  • Feiyu Zhao,
  • Yuxuan Tang,
  • Shuaidi He,
  • Xinghao Chen,
  • Qixiang Guo,
  • Minchao Zhang

摘要

The demands of increasing traffic congestion and urbanization have made the development of an intelligent transportation system (ITS) based on smart perception an inevitable trend. In recent years, based on the close integration of computer vision and UAV platforms, the intelligent detection technology of traffic targets from the perspective of UAVs has created an efficient and flexible perception network for the intelligent transportation system. In order to solve problems such as limited resolution, lack of texture, and complicated interference that arise while recognizing small traffic targets from the high-altitude perspective of drones, this study presented an effective end-to-end detector OF-DETR based on the RT-DETR algorithm. First, to better handle complicated background interference in high-altitude detection tasks, we created a Multiscale Global Feature Extraction Module (MGFEM) that enhanced the model’s capacity to model global contextual information. Second, we proposed the Cross-Scale Fusion Encoder, which reduced feature conflicts induced by cross-layer feature fusion and enhanced the representational capacity of detailed information, allowing the decoder to absorb richer feature inputs. In the loss function design, the FMPD-IoU loss function is proposed to direct the model to pay more attention to difficult samples and optimize bounding box regression, while MPDIoU-Aware Query Selection is proposed to offer more precise queries for the decoder. Our approach establishes a strong day-night visual perception foundation for air-ground collaborative intelligent transportation systems, outperforming RT-DETR by 3.36% during the day and 5.19% during the night on mAP50.