Scholarlink

<p>Underpinning smart monitoring, intelligent traffic monitoring systems, and public safety ecosystems, high-precision real-time crowd detection demands robustness against severe occlusion and drastic scale variation. CRD-YOLO, evolving the YOLOv11 framework, rectifies detection failures distinct to dense scenes. Central to this architecture, a Multi-Scale Partial Feature Extractor (MS-PFE) backbone dismantles computational redundancy, employing the Partial Feature Extractor (PFE) to synthesize multi-scale features with negligible overhead. Coupling an efficient RepGFPN structure with content-aware DySample upsampling, the Reparameterized Dynamic Feature Pyramid Network (RD-FPN) fortifies feature representation for minute, blurred targets, curbing information attrition inherent to upsampling processes. The architecture incorporates a Dynamic Head (DyHead), orchestrating a triple-attention mechanism across scale, spatial, and task dimensions to synchronize detection features. Disentangling feature representations of highly overlapping instances, this mechanism optimizes localization fidelity under severe occlusion. Empirical validations on the CrowdHuman and WiderPerson benchmarks substantiate the architecture’s dominance over contemporary state-of-the-art paradigms. Specifically, on the CrowdHuman dataset, the model attains 80.04% mAP@0.5 and 53.38% mAP@0.5-0.95, outperforming the baseline by 2.91% and 4.01%, respectively. Similarly, on WiderPerson, it attains 63.98% mAP@0.5 and 40.18% mAP@0.5-0.95. The source code for the experiments is openly available at <a href="https://github.com/YY-258-bit/CRD-YOLO">https://github.com/YY-258-bit/CRD-YOLO</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

CRD-YOLO: a high-accuracy real-time crowded pedestrian detection algorithm

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Zhaohui Hu,
  • Wuyang Niu,
  • Shuai Mo

摘要

Underpinning smart monitoring, intelligent traffic monitoring systems, and public safety ecosystems, high-precision real-time crowd detection demands robustness against severe occlusion and drastic scale variation. CRD-YOLO, evolving the YOLOv11 framework, rectifies detection failures distinct to dense scenes. Central to this architecture, a Multi-Scale Partial Feature Extractor (MS-PFE) backbone dismantles computational redundancy, employing the Partial Feature Extractor (PFE) to synthesize multi-scale features with negligible overhead. Coupling an efficient RepGFPN structure with content-aware DySample upsampling, the Reparameterized Dynamic Feature Pyramid Network (RD-FPN) fortifies feature representation for minute, blurred targets, curbing information attrition inherent to upsampling processes. The architecture incorporates a Dynamic Head (DyHead), orchestrating a triple-attention mechanism across scale, spatial, and task dimensions to synchronize detection features. Disentangling feature representations of highly overlapping instances, this mechanism optimizes localization fidelity under severe occlusion. Empirical validations on the CrowdHuman and WiderPerson benchmarks substantiate the architecture’s dominance over contemporary state-of-the-art paradigms. Specifically, on the CrowdHuman dataset, the model attains 80.04% mAP@0.5 and 53.38% mAP@0.5-0.95, outperforming the baseline by 2.91% and 4.01%, respectively. Similarly, on WiderPerson, it attains 63.98% mAP@0.5 and 40.18% mAP@0.5-0.95. The source code for the experiments is openly available at https://github.com/YY-258-bit/CRD-YOLO.