{
  "paper_id": "cb5537d32b79f95a7134c983c56fa558",
  "paper_title": "Top view multiple people tracking by detection using deep SORT and YOLOv3 with transfer learning: within 5G infrastructure",
  "paper_authors": [
    "Imran Ahmed",
    "Misbah Ahmad",
    "Awais Ahmad",
    "Gwanggil Jeon"
  ],
  "publication_date": "2020-10-01",
  "publication_venue": "Journal of Ambient Intelligence and Humanized Computing",
  "abstract": "Nowadays, 5G profoundly impacts video surveillance and monitoring services by processing video streams at high-speed with high-reliability, high bandwidth, and secure network connectivity. It also enhances artificial intelligence, machine learning, and deep learning techniques, which require intense processing to deliver near-real-time solutions. In video surveillance, person tracking is a crucial task due to the deformable nature of the human body, various environmental components such as occlusion, illumination, and background conditions, specifically, from a top view perspective where the person's visual appearance is significantly different from a frontal or side view. In this work, multiple people tracking framework is presented, which uses 5G infrastructure. A top view perspective is used, which offers broad coverage of the scene or field of view. To perform a person tracking deep learning-based tracking by detection framework is proposed, which includes detection by YOLOv3 and tracking by Deep SORT algorithm. Although the model is pre-trained using the frontal view images, even then, it gives good detection results. In order to further enhance the accuracy of the detection model, the transfer learning approach is adopted. In this way, a detection model takes advantage of a pre-trained model appended with an additional trained layer using top view data set. To evaluate the performance, experiments are carried out on different top view video sequences. Experimental results reveal that transfer learning improves the overall performance, detection accuracy, and reduces false positives. The deep learning detection model YOLOv3 achieves detection accuracy of 92% with a pre-trained model without transfer learning and 95% with transfer learning. The tracking algorithm Deep SORT also achieves excellent results with a tracking accuracy of 96%.",
  "keywords": [
    "5G",
    "Deep learning",
    "YOLOv3",
    "Deep SORT",
    "Transfer learning",
    "Person detection and tracking",
    "Top view"
  ],
  "paper_institutions": [
    "Institute of Management Sciences, Pakistan",
    "Department of Computer Science, Bahria University, Islamabad, Pakistan",
    "Department of Embedded Systems Engineering, Incheon National University, South Korea",
    "Department of Electronic Engineering, Xidian University, China"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "ark-code-latest",
  "prompt_template": "simple",
  "extracted_at": "2026-08-16T19:46:12.564295Z",
  "template_tag": "simple",
  "summary": "### 摘要\n本文针对在5G基础设施下俯视视角的多行人跟踪问题，提出了一种基于检测的深度学习跟踪框架。由于俯视视角下行人外观与正面或侧面差异显著，且传统监控中遮挡问题严重，采用单俯视摄像头可有效缓解遮挡并降低部署成本。该框架使用YOLOv3进行行人检测，并使用Deep SORT进行多目标跟踪；模型先在COCO数据集上预训练，再采用迁移学习在自建的俯视行人数据集上额外训练，并通过合并权重文件提升检测性能。实验在室内校园环境采集的多个俯视视频序列上进行，包含2至20人，使用精确率、召回率、准确率以及MOTA和MOTP等指标评估。结果表明，YOLOv3检测准确率从无迁移学习的92%提升至有迁移学习的95%，Deep SORT跟踪准确率达96%，验证了迁移学习对俯视场景检测的有效性。未来工作将扩展至多类别俯视目标检测，并探索更丰富的场景。核心公式为检测置信度Conf(person)=Pr(person)×IOU(Pred,Truth)，通过预测框与真实框的IOU以及损失函数联合优化检测结果。\n\n### 关键要点\n- 针对俯视视角下行人外观变化大、遮挡问题严重，提出基于5G基础设施的多人跟踪框架。\n- 采用单俯视摄像头获取视频流，通过5G网络传输至服务器，再利用深度学习模型进行检测和跟踪。\n- 检测采用YOLOv3，预训练于COCO数据集，模型原本针对正面或侧面视角训练。\n- 为提升俯视检测精度，采用迁移学习：使用自建俯视行人数据集（1000个图像块，10,000帧）额外训练YOLOv3，并将新权重与预训练权重合并。\n- 跟踪采用Deep SORT算法，结合卡尔曼滤波和外观描述符，通过马氏距离和余弦距离构建代价矩阵进行数据关联。\n- 实验视频序列在室内校园环境采集，使用Point Gray FlyCap2相机和广角镜头，安装高度约4米，分辨率644×482，帧率20fps。\n- 数据分集：训练集70%，测试集30%；训练批次大小64，迭代40轮。\n- 无迁移学习时，YOLOv3检测准确率为92%，有迁移学习时提升至95%。\n- Deep SORT跟踪的MOTA为92%（无迁移）和96%（有迁移），MOTP为90%和95%。\n- 实验结果表明迁移学习减少了误检（假阳性），并在遮挡和光照变化条件下仍能稳定跟踪行人。\n- 模型在俯视场景中可区分紧密接触的行人，并能正确分配跟踪ID，但在某些情况下仍存在ID交换或漏检。\n- 该工作是首次在5G基础设施下对俯视多人数据集使用迁移学习训练YOLOv3并耦合Deep SORT进行跟踪的尝试。"
}