{
  "paper_id": "b440d5c892180565eb0f355c09c1d4ed",
  "paper_title": "Automatic Labeled LiDAR Data Generation based on Precise Human Model",
  "paper_authors": [
    "Wonjik Kim",
    "Masayuki Tanaka",
    "Masatoshi Okutomi",
    "Yoko Sasaki"
  ],
  "publication_date": "",
  "publication_venue": "2019 International Conference on Robotics and Automation (ICRA)",
  "abstract": "Following improvements in deep neural networks, state-of-the-art networks have been proposed for human recognition using point clouds captured by LiDAR. However, the performance of these networks strongly depends on the training data. An issue with collecting training data is labeling. Labeling by humans is necessary to obtain the ground truth label; however, labeling requires huge costs. Therefore, we propose an automatic labeled data generation pipeline, for which we can change any parameters or data generation environments. Our approach uses a human model named Dhaiba and a background of Miraikan and consequently generated realistic artificial data. We present 500k+ data generated by the proposed pipeline. This paper also describes the specification of the pipeline and data details with evaluations of various approaches.",
  "keywords": [
    "LiDAR",
    "data generation",
    "human segmentation",
    "point cloud",
    "deep learning"
  ],
  "paper_institutions": [],
  "output_language": "zh",
  "provider": "doubao",
  "model": "ark-code-latest",
  "prompt_template": "simple",
  "extracted_at": "2026-08-16T19:46:44.139262Z",
  "template_tag": "simple",
  "summary": "### 摘要\n本文针对LiDAR点云人体识别任务中训练数据标注成本高昂的问题，提出了一种自动生成带标注深度图的流水线。该流水线基于精确数字人体模型Dhaiba和从东京科学未来馆采集的背景深度图，通过控制人体身高、体重、行走姿态及LiDAR参数等，自动合成包含深度、xyz坐标和人体标签的数据，无需人工标注。研究者利用该流水线生成了超过50万张深度图，并使用47K生成数据训练了FCN、U-net、FCDN和Pointnet四种分割网络，在0.1K人工标注的真实数据上评估。实验表明，增加训练数据量可提升平均精度、召回率和威胁分数（IoU）；基于精确行走模型的数据优于站立模型；未参与训练的身高体重组合表现与训练组合相当，表明泛化性良好。背景差异会导致性能下降，未来需扩展多种背景。核心机制是通过已知LiDAR与人体模型的三维位置，利用像素级最小深度合成将人体深度图与背景深度图融合，并自动获得标签。\n\n### 关键要点\n- 提出全自动标注LiDAR点云数据生成流水线，包含背景采集、人体模型构建和深度图合成三步，无需人工标注。\n- 采用数字人体模型Dhaiba，可生成15种身高体重组合与230帧行走动作，共3450种不同行走模型。\n- 背景数据在东京科学未来馆采集，使用Velodyne HDL-32E LiDAR，共2363张背景深度图。\n- 生成数据包含深度图、xyz坐标和标签，以HDF5格式存储，并附有详细xml信息。\n- 实验使用47K生成数据训练，0.1K真实标注数据验证，模型距离分布集中在0-5m（74.45%）。\n- 增加训练数据量（0.1K到47K）可提升平均精度、召回率和威胁分数，但精度略有下降。\n- 行走模型比站立模型平均精度更高（0.6922 vs 0.6058），证明精确模型的有效性。\n- 未训练的身高体重组合（如1000mm-12kg）性能与训练组合相当，表明网络对体型的泛化能力。\n- 背景差异显著影响性能：使用不同背景训练时威胁分数降至0.1018，而相同背景为0.3919。\n- 四种网络中，FCN在深度输入下平均精度最高（生成数据0.7529，真实数据0.7095）。\n- FCDN在深度+xyz输入下召回率最高（生成数据0.7707，真实数据0.7848）。\n- 生成测试数据的性能可预测真实数据表现，但Pointnet在真实数据上召回率较低（0.0865）。"
}