Scholarlink

<p>With recent advancements in vision intelligence, pedestrian detection in autonomous driving has become a critical research focus within computer vision. Dense pedestrian scenarios present significant challenges from multi-scale variations and occlusions. Traditional detection methods can be used for pedestrian detection in ordinary scenarios, but they face challenges, such as high computational complexity and overly sophisticated models, that prevent effective deployment on mobile devices like in-vehicle cameras, along with unsatisfactory detection accuracy under multi-scale pedestrian scenarios and heavy occlusion conditions. To address these challenges, this paper proposes YOLO-GSD, an improved lightweight real-time pedestrian detection algorithm based on YOLOv8. The algorithm first introduces a dedicated detection layer specifically designed for small-scale targets. It then incorporates lightweight Ghost convolution and designs a DG-C2f module by integrating Ghost convolution and Dynamic Convolution, aiming to reduce computational complexity while enhancing the algorithm’s multi-scale feature fusion capability. Additionally, it employs the ultra-lightweight DySample upsampler for efficient feature reconstruction and integrates the SEAM attention mechanism to improve occlusion-aware detection. Finally, WIoUv3 is used to replace the CIoU loss function, which improves the generalization ability and overall performance of the algorithm. Experimental results demonstrate mAP@0.5 scores of 90.7% on the WiderPerson dataset (1.4% higher than the baseline) and 86.5% on the CrowdHuman dataset (2.2% improvement). The algorithm’s parameter count is reduced to 6.24 M, its FLOPs are lowered to 22.7 G, and its FPS is increased to 106.6. In addition, a homogeneous training comparison was conducted on the small-object dataset RSOD, demonstrating the advantages of YOLO-GSD in small-object detection. Therefore, the YOLO-GSD algorithm proposed in this paper is suitable for real-time pedestrian detection in multi-scale and occlusion scenarios on mobile platforms with limited computational resources.</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

YOLO-GSD: a real-time pedestrian detection algorithm based on YOLOv8 in dense environments

  • 期刊论文

  • 发表日期：

    2025年9月25日

查看全文
  • Zuhao Zhang,
  • Weiwei Li,
  • Lin Luo

摘要

With recent advancements in vision intelligence, pedestrian detection in autonomous driving has become a critical research focus within computer vision. Dense pedestrian scenarios present significant challenges from multi-scale variations and occlusions. Traditional detection methods can be used for pedestrian detection in ordinary scenarios, but they face challenges, such as high computational complexity and overly sophisticated models, that prevent effective deployment on mobile devices like in-vehicle cameras, along with unsatisfactory detection accuracy under multi-scale pedestrian scenarios and heavy occlusion conditions. To address these challenges, this paper proposes YOLO-GSD, an improved lightweight real-time pedestrian detection algorithm based on YOLOv8. The algorithm first introduces a dedicated detection layer specifically designed for small-scale targets. It then incorporates lightweight Ghost convolution and designs a DG-C2f module by integrating Ghost convolution and Dynamic Convolution, aiming to reduce computational complexity while enhancing the algorithm’s multi-scale feature fusion capability. Additionally, it employs the ultra-lightweight DySample upsampler for efficient feature reconstruction and integrates the SEAM attention mechanism to improve occlusion-aware detection. Finally, WIoUv3 is used to replace the CIoU loss function, which improves the generalization ability and overall performance of the algorithm. Experimental results demonstrate mAP@0.5 scores of 90.7% on the WiderPerson dataset (1.4% higher than the baseline) and 86.5% on the CrowdHuman dataset (2.2% improvement). The algorithm’s parameter count is reduced to 6.24 M, its FLOPs are lowered to 22.7 G, and its FPS is increased to 106.6. In addition, a homogeneous training comparison was conducted on the small-object dataset RSOD, demonstrating the advantages of YOLO-GSD in small-object detection. Therefore, the YOLO-GSD algorithm proposed in this paper is suitable for real-time pedestrian detection in multi-scale and occlusion scenarios on mobile platforms with limited computational resources.