Scholarlink

<p>AAccurate and efficient detection of small or dim traffic objects remains a major challenge in intelligent transportation systems. To address these issues, this paper proposes WS-DETR, a real-time DETR-based detection framework designed to enhance small-object perception and reduce false detections under low-light conditions. The model integrates two key components. First, the WM-Dual Block utilizes wavelet decomposition as a structural prior for the Mamba state-space model, guiding the scanning mechanism to explicitly capture complementary high-frequency details and low-frequency semantics. Second, the SQ-ACA mechanism employs shared-query axial cross-attention to ensure feature consistency across spatial directions. Extensive experiments on three benchmark datasets–M3FD, BDD100K, and UA-DETRAC–demonstrate that WS-DETR achieves superior performance compared with state-of-the-art YOLO and DETR variants, particularly in small-object and low-light scenarios. The source code will be publicly available at&#xa0;<a href="https://github.com/ZehuaChenLab/WSDETR">https://github.com/ZehuaChenLab/WSDETR</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

WS-DETR: A Wavelet-mamba and shared query attention enhanced DETR for traffic object detection

  • 期刊论文

  • 发表日期：

    2026年2月19日

查看全文
  • Lianghao Xu,
  • ShuHui Liu,
  • Zehua Chen,
  • Shan Wang

摘要

AAccurate and efficient detection of small or dim traffic objects remains a major challenge in intelligent transportation systems. To address these issues, this paper proposes WS-DETR, a real-time DETR-based detection framework designed to enhance small-object perception and reduce false detections under low-light conditions. The model integrates two key components. First, the WM-Dual Block utilizes wavelet decomposition as a structural prior for the Mamba state-space model, guiding the scanning mechanism to explicitly capture complementary high-frequency details and low-frequency semantics. Second, the SQ-ACA mechanism employs shared-query axial cross-attention to ensure feature consistency across spatial directions. Extensive experiments on three benchmark datasets–M3FD, BDD100K, and UA-DETRAC–demonstrate that WS-DETR achieves superior performance compared with state-of-the-art YOLO and DETR variants, particularly in small-object and low-light scenarios. The source code will be publicly available at https://github.com/ZehuaChenLab/WSDETR.