Scholarlink

<p>Traffic environment perception constitutes a critical real-time computing task for intelligent driving systems, which must support instantaneous decision-making under strict latency constraints. However, existing traffic object detection methods often lack robustness in complex adverse weather conditions and tend to rely on computationally intensive architectures. Meanwhile, collecting and annotating large-scale traffic images under adverse weather conditions is costly and challenging. To tackle the above challenges, we propose a progressive cross-domain adaptive YOLO with image translation (TPCDA-YOLO), designed for real-time high-performance object detection in traffic environments. First, we build a dual-branch image translation network (DBITNet) to generate inter-domain interpolation images, constructing auxiliary domains to remedy domain discrepancy. Using YOLOv10 as the backbone network for detection, we propose a pixel-level feature adaptation (PFA) module to effectively alleviate confusion about object details through local central region perception under adverse weather conditions. Additionally, we incorporate a multi-scale image-level feature adaptation (MS-IFA) module to facilitate progressively cross-domain feature alignment, transitioning from the normal to the adverse weather domain. To further enhance the high discriminability of the label space across different abstraction levels, intra-domain consistent regularization (IDCR) module is proposed to reduce the domain category joint distribution matching offset, thereby identifying challenging alignment cases. Extensive experiments on multiple benchmark cross-domain traffic datasets demonstrate that our proposed TPCDA-YOLO achieves improved detection performance under adverse weather conditions compared to existing domain adaptation methods, while maintaining high inference efficiency, thereby satisfying the stringent real-time requirements of intelligent driving systems.</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

TPCDA-YOLO: progressive cross-domain adaptive YOLO with image translation for adverse weather traffic scenarios

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Yanqi Sun,
  • Yong Duan,
  • Lin Zhang,
  • Cuicui Li,
  • Jianhua Li

摘要

Traffic environment perception constitutes a critical real-time computing task for intelligent driving systems, which must support instantaneous decision-making under strict latency constraints. However, existing traffic object detection methods often lack robustness in complex adverse weather conditions and tend to rely on computationally intensive architectures. Meanwhile, collecting and annotating large-scale traffic images under adverse weather conditions is costly and challenging. To tackle the above challenges, we propose a progressive cross-domain adaptive YOLO with image translation (TPCDA-YOLO), designed for real-time high-performance object detection in traffic environments. First, we build a dual-branch image translation network (DBITNet) to generate inter-domain interpolation images, constructing auxiliary domains to remedy domain discrepancy. Using YOLOv10 as the backbone network for detection, we propose a pixel-level feature adaptation (PFA) module to effectively alleviate confusion about object details through local central region perception under adverse weather conditions. Additionally, we incorporate a multi-scale image-level feature adaptation (MS-IFA) module to facilitate progressively cross-domain feature alignment, transitioning from the normal to the adverse weather domain. To further enhance the high discriminability of the label space across different abstraction levels, intra-domain consistent regularization (IDCR) module is proposed to reduce the domain category joint distribution matching offset, thereby identifying challenging alignment cases. Extensive experiments on multiple benchmark cross-domain traffic datasets demonstrate that our proposed TPCDA-YOLO achieves improved detection performance under adverse weather conditions compared to existing domain adaptation methods, while maintaining high inference efficiency, thereby satisfying the stringent real-time requirements of intelligent driving systems.