Scholarlink

<p>In dense pedestrian detection tasks, due to the high proportion of small-scale pedestrians, dense spatial distribution, and complex background interference, existing detection models are prone to false positives and missed detections when handling occlusion and scale variations. To address these challenges, this paper proposes FEOD-YOLOv11, a dense pedestrian detection algorithm based on YOLOv11, which features fine-grained feature enhancement and occlusion-aware for complex scenes. Firstly, to effectively capture fine-grained features of small-scale pedestrians in densely crowded scenes, we integrate the MetaFormer architecture with deformable convolution and depthwise separable convolution to strengthen the backbone’s fine-grained feature extraction capability. Secondly, we design a C2PSA-TJA module to enhance deep feature representation, thereby strengthening the features of the visible parts of occluded pedestrians and suppressing interference from complex backgrounds. Thirdly, we introduce a small-scale pedestrian detection head specifically for small-scale pedestrians, which combines high-resolution features with deep semantic information to improve detection performance. Finally, we design a novel loss function called Wise-Inner-ECIoU, which suppresses gradient noise from low-quality samples and enhances overall detection accuracy. Experiments on the CrowdHuman dataset demonstrate that FEOD-YOLOv11 outperforms the baseline model, achieving significant improvements in both <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\textit{AP}_\text {50}\)</EquationSource> </InlineEquation> and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\textit{MR}^\text {-2}\)</EquationSource> </InlineEquation>. Moreover, the model exhibits strong generalization capability on the WiderPerson dataset in complex scenes. These results confirm the effectiveness and robustness of the proposed method for dense pedestrian detection under challenging conditions.</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

Fine-grained feature enhancement and occlusion-aware dense pedestrian detection in complex scenes

  • 期刊论文

  • 发表日期：

    2025年12月4日

查看全文
  • Mingxing Li,
  • Shucheng Huang

摘要

In dense pedestrian detection tasks, due to the high proportion of small-scale pedestrians, dense spatial distribution, and complex background interference, existing detection models are prone to false positives and missed detections when handling occlusion and scale variations. To address these challenges, this paper proposes FEOD-YOLOv11, a dense pedestrian detection algorithm based on YOLOv11, which features fine-grained feature enhancement and occlusion-aware for complex scenes. Firstly, to effectively capture fine-grained features of small-scale pedestrians in densely crowded scenes, we integrate the MetaFormer architecture with deformable convolution and depthwise separable convolution to strengthen the backbone’s fine-grained feature extraction capability. Secondly, we design a C2PSA-TJA module to enhance deep feature representation, thereby strengthening the features of the visible parts of occluded pedestrians and suppressing interference from complex backgrounds. Thirdly, we introduce a small-scale pedestrian detection head specifically for small-scale pedestrians, which combines high-resolution features with deep semantic information to improve detection performance. Finally, we design a novel loss function called Wise-Inner-ECIoU, which suppresses gradient noise from low-quality samples and enhances overall detection accuracy. Experiments on the CrowdHuman dataset demonstrate that FEOD-YOLOv11 outperforms the baseline model, achieving significant improvements in both \(\textit{AP}_\text {50}\) and \(\textit{MR}^\text {-2}\) . Moreover, the model exhibits strong generalization capability on the WiderPerson dataset in complex scenes. These results confirm the effectiveness and robustness of the proposed method for dense pedestrian detection under challenging conditions.