Scholarlink

<p>Robust pedestrian detection in unconstrained environments remains challenging due to adverse weather conditions, occlusions, scale variations, and visually confusing background objects. This paper presents CamPedV2, a large-scale real-world pedestrian detection dataset designed to support systematic analysis of these challenges. CamPedV2 comprises over 114k images with more than 545k pedestrian annotations, collected under diverse lighting conditions, natural fog and rain, heavy occlusion, varied viewpoints, and the presence of person-like distractors. Beyond dataset construction, we provide an in-depth statistical characterization of pedestrian scale, density, and structural scene complexity, highlighting differences from existing benchmarks. Extensive experiments using both CNN-based and transformer-based detectors, including YOLOv7–YOLOv12, DETR, and RT-DETR, demonstrate that models trained on CamPedV2 achieve improved robustness in challenging scenarios and stronger cross-dataset generalization. Additional ablation studies analyze the contribution of weather conditions and person-like object filtering. We further discuss the domain gaps between campus and urban scenes, outlining limitations and future research directions. CamPedV2 is publicly released to facilitate research on robust pedestrian detection in real-world conditions. We have open-sourced this data set on GitHub. You can access it at <a href="https://github.com/RahulRaman2/CamPedV2">https://github.com/RahulRaman2/CamPedV2</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

CamPedV2: a diverse real-time pedestrian detection dataset for challenging environmental conditions

  • 期刊论文

  • 发表日期：

    2026年6月21日

查看全文
  • V. S. Sukesh Babu,
  • Rahul Raman

摘要

Robust pedestrian detection in unconstrained environments remains challenging due to adverse weather conditions, occlusions, scale variations, and visually confusing background objects. This paper presents CamPedV2, a large-scale real-world pedestrian detection dataset designed to support systematic analysis of these challenges. CamPedV2 comprises over 114k images with more than 545k pedestrian annotations, collected under diverse lighting conditions, natural fog and rain, heavy occlusion, varied viewpoints, and the presence of person-like distractors. Beyond dataset construction, we provide an in-depth statistical characterization of pedestrian scale, density, and structural scene complexity, highlighting differences from existing benchmarks. Extensive experiments using both CNN-based and transformer-based detectors, including YOLOv7–YOLOv12, DETR, and RT-DETR, demonstrate that models trained on CamPedV2 achieve improved robustness in challenging scenarios and stronger cross-dataset generalization. Additional ablation studies analyze the contribution of weather conditions and person-like object filtering. We further discuss the domain gaps between campus and urban scenes, outlining limitations and future research directions. CamPedV2 is publicly released to facilitate research on robust pedestrian detection in real-world conditions. We have open-sourced this data set on GitHub. You can access it at https://github.com/RahulRaman2/CamPedV2.