Scholarlink

<p>Pedestrian detection is a critical component of Intelligent Transportation Systems (ITS). Although state-of-the-art detectors perform well under normal traffic conditions, they struggle in congested, panicked, and irregular motion scenarios, where pedestrians often exhibit frequent occlusions, diverse postures, and unstable movements. Accurately detecting pedestrians in these challenging environments is difficult, as it requires not only maintaining high confidence for each detected pedestrian but also precisely predicting bounding box parameters for pedestrians of varying scales within the same scene. To this end, we follow the divide-and-conquer strategy of mainstream detectors and propose a Deep Aggregation Layer (DLA) network for cross-stage feature enhancement, called MFEDet (Multi-scale Feature Enhancement Detector). Extensive experiments on 2 popular and challenging pedestrian datasets, HiEve and CUHK Occlusion, demonstrate that MFEDet achieves optimal balance between speed and accuracy compared to baseline models. Our code and data are available at <a href="https://github.com/YanJieWen/DLA_YOLOv3-for-complex-events">https://github.com/YanJieWen/DLA_YOLOv3-for-complex-events</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

Multi-scale feature enhancement for pedestrian detection in complex events: the elephant in the room

  • 期刊论文

  • 发表日期：

    2025年10月31日

查看全文
  • Yanjie Wen,
  • Ping Xu,
  • Zhihong Li

摘要

Pedestrian detection is a critical component of Intelligent Transportation Systems (ITS). Although state-of-the-art detectors perform well under normal traffic conditions, they struggle in congested, panicked, and irregular motion scenarios, where pedestrians often exhibit frequent occlusions, diverse postures, and unstable movements. Accurately detecting pedestrians in these challenging environments is difficult, as it requires not only maintaining high confidence for each detected pedestrian but also precisely predicting bounding box parameters for pedestrians of varying scales within the same scene. To this end, we follow the divide-and-conquer strategy of mainstream detectors and propose a Deep Aggregation Layer (DLA) network for cross-stage feature enhancement, called MFEDet (Multi-scale Feature Enhancement Detector). Extensive experiments on 2 popular and challenging pedestrian datasets, HiEve and CUHK Occlusion, demonstrate that MFEDet achieves optimal balance between speed and accuracy compared to baseline models. Our code and data are available at https://github.com/YanJieWen/DLA_YOLOv3-for-complex-events.