{
  "paper_id": "3d98f931368644f310ec3648fa83cbdb",
  "paper_title": "CNN-based Human Detection Using a 3D LiDAR onboard a UAV",
  "paper_authors": [
    "Jack N.C. Hayton",
    "Tiago Barros",
    "Cristiano Premebida",
    "Matthew J. Coombes",
    "Urbano J. Nunes"
  ],
  "publication_date": "April 15-16, 2020",
  "publication_venue": "2020 IEEE International Conference on Autonomous Robot Systems and Competitions (ICARSC)",
  "abstract": "This paper addresses the problem of detecting humans in a point cloud taken with a 3D-LiDAR onboard a UAV. The potential use cases of this technology are numerous, examples include security and surveillance, disaster relief and search and rescue operations. In this paper, a CNN-based approach is proposed which is able to analyse point clouds returned by a 3D LiDAR sensor in such a way that it can detect humans. The algorithm described here consists of 3 main components: data pre-processing, post-processing, and human classification. In this paper, objects were assigned to one of two classes: human and non-human. The classification was performed by projecting the 3D point cloud onto a series of 2D planes using occupancy grid mapping. This creates a set of silhouettes of the object corresponding to the top, front and side views. Classification is achieved by supervised CNNs using single-view and multi-view (3 channels) images patches.",
  "keywords": [
    "UAV",
    "3D LiDAR",
    "human detection",
    "CNN",
    "occupancy grid",
    "point cloud",
    "object classification"
  ],
  "paper_institutions": [
    "Centre for Autonomous and Cyberphysical Systems, Cranfield University",
    "Institute of Systems and Robotics, University of Coimbra"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "ark-code-latest",
  "prompt_template": "simple",
  "extracted_at": "2026-08-16T19:46:38.736984Z",
  "template_tag": "simple",
  "summary": "### 摘要\n本文研究利用无人机搭载的3D LiDAR传感器采集的点云数据进行人体检测的问题。该技术可用于安全监控、灾害救援和搜救等场景。提出了一种基于CNN的检测方法，包含数据预处理、后处理和人分类三个主要模块。预处理包括同步、地面检测、对齐和ROI生成；采用层次聚类（基于欧氏距离）生成对象假设；通过将3D点云投影到三个正交平面（顶部、前视、侧视）并利用占用栅格映射生成2D图像，将投影图像作为CNN输入，采用单视图和多视图（三通道）两种方式训练分类器。实验使用自建的无人机LiDAR数据集，包含5名成人和10种非人物体，数据增强后共3124个样本，采用4折交叉验证评估模型性能。结果显示，多视图（TFS）增强模型获得了98.8%的准确率和99.9%的AUC，优于单视图模型。方法在简单平地场景下有效，但聚类环节在目标距离较近时成功率较低，未来可改进聚类算法并引入反馈机制提高鲁棒性。\n\n### 关键要点\n- 本文是首个利用现成的机载3D LiDAR（Velodyne VLP-16）在无人机上进行人体检测的研究，提供了真实世界数据集。\n- 算法流程包括预处理（同步、地面检测、对齐）、聚类生成对象假设和CNN分类三个步骤。\n- 地面检测采用RANSAC平面拟合，并通过两次旋转对齐，平均误差从4.6度降至0.45度。\n- 使用层次聚类（kd-tree + 欧氏距离）进行对象分割，设置距离阈值2米，在单目标场景成功率96%，多目标时性能下降。\n- 采用占用栅格映射将3D点云投影到三个正交平面（T、F、S），生成90×90分辨率的二进制图像。\n- 对比了不同网格分辨率（10、25、50 cells），发现25 cells效果最佳。\n- CNN架构包含卷积、批归一化、ReLU、池化和全连接层，输入为单通道或三通道图像。\n- 数据集包含225个非人样本和285个人类样本，数据增强（绕三个轴旋转）后样本总量增至3124个。\n- 使用4折交叉验证评估模型，多视图（TFS）增强CNN达到98.8%准确率和98.9% F-score，优于单视图模型。\n- 顶部视图（T）性能最差（准确率78.2%），侧视图（S）增强后准确率略降（95.5%），但非增强模型侧视图准确率更高。\n- 聚类模块是算法的主要瓶颈，在目标间距小于1米时成功率仅57%，需改进。\n- 算法假设平坦地面，实际应用中需要处理不平坦和多层级地面，未来可引入反馈循环和更复杂的检测方法（如Complex-YOLO）。",
  "available_templates": [
    "deep_read",
    "simple"
  ]
}