{
  "paper_id": "0a9d715ef7e2707bb888a05a4f195b8e",
  "paper_title": "Where are we with Human Pose Estimation in Real-World Surveillance?",
  "paper_authors": [
    "Mickael Cormier",
    "Andreas Specker",
    "Jürgen Beyerer",
    "Aris Clepe"
  ],
  "publication_date": "",
  "publication_venue": "2022 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW)",
  "abstract": "The rapidly increasing number of surveillance cameras offers a variety of opportunities for intelligent video analytics to improve public safety. Among many others, the automatic recognition of suspicious and violent behavior poses a key task. To preserve personal privacy, prevent ethnic bias, and reduce complexity, most approaches first extract the pose or skeleton of persons and subsequently perform activity recognition. However, current literature mainly focuses on research datasets and does not consider real-world challenges and requirements of human pose estimation. We close this gap by analyzing these challenges, such as inadequate data and the need for real-time processing, and proposing a framework for human pose estimation in uncontrolled crowded surveillance scenarios. Our system integrates mitigation measures as well as a tracking component to incorporate temporal information. Finally, we provide a detailed quantitative and qualitative analysis on both a scientific and a real-world dataset to highlight improvements and remaining obstacles towards robust real-world human pose estimation in uncooperative scenarios.",
  "keywords": [
    "Human Pose Estimation",
    "Surveillance",
    "Real-world",
    "Crowd",
    "Tracking",
    "Temporal consistency"
  ],
  "paper_institutions": [
    "Fraunhofer IOSB",
    "Fraunhofer Center for Machine Learning",
    "Karlsruhe Institute of Technology (KIT)"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "ark-code-latest",
  "prompt_template": "simple",
  "extracted_at": "2026-08-16T19:46:30.748469Z",
  "template_tag": "simple",
  "summary": "### 摘要\n本文针对真实监控场景中的人体姿态估计（HPE）问题进行了系统分析。指出了现有研究多聚焦于科研数据集，未考虑真实世界的挑战，如数据不足、实时处理需求、遮挡、视角变化等。提出了一套适用于非受控拥挤监控场景的HPE框架，包含人员检测、视觉跟踪和姿态估计三个主要模块，并融合了时间一致性策略。在PRW、PANDA以及自建的RWS-CPE数据集上进行了定量和定性评估。结果表明，基于YOLOv4-Panda的检测器在真实监控数据上取得最佳性能；Transpose混合变换器模型在姿态估计上表现出更好的泛化能力，但在检测框不准确时性能下降明显。系统以7-10 FPS运行，验证了其在实际部署中的可行性，同时指出仍需大规模真实标注数据以缩小领域差距。\n### 关键要点\n- 真实监控场景中的人物检测和姿态估计面临多种挑战，包括遮挡、相机角度、光照变化、反光、阴影等。\n- 现有公开数据集（如COCO、MPII、PoseTrack18、CrowdPose）与监控域存在巨大差异，导致模型泛化能力不足。\n- 自建RWS-CPE数据集从11个摄像头选取100帧，标注4785个边界框和1894个姿态，物体大小分布不均，以小型和中等尺寸为主。\n- 系统采用YOLOv4作为人员检测器，并在PANDA数据集上微调，获得了在真实监控数据上更好的检测性能（AP 0.28 vs 其他模型）。\n- 视觉跟踪模块使用OSNet-x1.0实例归一化模型，结合数据增强（翻转、亮度和对比度调整、旋转），并在多个数据集（Market-1501、DukeMTMC-reID、PRAI-1581）上训练，实现了10 FPS的跟踪速度。\n- 为了保持时间一致性，采用环形缓冲保存最近5帧的特征，对缺失检测进行插值或填充，并过滤尺寸过小或无效的轨迹。\n- 姿态估计实验表明，Transpose-R和Transpose-H在COCO和PoseTrack18预训练后，在RWS-CPE上使用GT框时AP分别达到0.60和0.59，优于HRNet等模型，但面对预测框时AP降至0.49和0.44。\n- 模型推理速度普遍较慢，RTX 2060上最高仅7 FPS（ResNet-50），实际系统中通过FP16和批处理等技术提升，但仍有改进空间。\n- 失败案例分析显示，包含多个人员、婴儿车或自行车等干扰物时，模型容易出现关键点错位，且难以区分左右。\n- 作者指出，未来需要更多的真实标注数据和半监督/交互式标注方法，并可通过关键点异常检测来进一步提升时间一致性。",
  "available_templates": [
    "deep_read",
    "simple"
  ]
}