{
  "paper_id": "8d9fc1b2f8683a5425350b6cd8bd676a",
  "paper_title": "CNN-based People Detection in Voxel Space using Intensity Measurements and Point Cluster Flattening",
  "paper_authors": [
    "J. Dybedal",
    "G. Hovland"
  ],
  "publication_date": "",
  "publication_venue": "Modeling, Identification and Control: A Norwegian Research Bulletin",
  "abstract": "In this paper real-time people detection is demonstrated in a relatively large indoor industrial robot cell as well as in an outdoor environment. Six depth sensors mounted at the ceiling are used to generate a merged point cloud of the cell. The merged point cloud is segmented into clusters and flattened into gray-scale 2D images in the xy and xz planes. These images are then used as input to a classifier based on convolutional neural networks (CNNs). The final output is the 3D position $(x, y, z)$ and bounding box representing the human. The system is able to detect and track multiple humans in real-time, both indoors and outdoors. The positional accuracy of the proposed method has been verified against several ground truth positions, and was found to be within the point-cloud voxel-size used, i.e. 0.04 m. Tests on outdoor datasets yielded a detection recall of 76.9% and an F1 score of 0.87.",
  "keywords": [
    "Human detection",
    "indoor",
    "outdoor",
    "point clouds",
    "voxels",
    "flattening",
    "CNN"
  ],
  "paper_institutions": [
    "Department of Engineering Sciences, University of Agder, N-4898 Grimstad, Norway"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "ark-code-latest",
  "prompt_template": "simple",
  "extracted_at": "2026-08-16T19:46:35.301542Z",
  "template_tag": "simple",
  "summary": "### 摘要\n\n本文针对仅依赖深度信息（无RGB图像）进行人员检测的问题，提出了一种基于点簇展平和卷积神经网络（CNN）场景分类的方法。相关研究多依赖于激光雷达或RGB-D数据，而在油井平台红色区域等需要冗余的场合，深度传感器信号可能成为唯一可用信息。该方法利用六个固定安装的Kinect V2深度传感器生成合并点云，经体素化、聚类和维度约束后，将候选点簇在xz和yz平面展平为灰度图像，并输入到一个简单的CNN分类器（含三个卷积块）中进行二分类（人/非人）。核心机制是通过累加体素强度生成像素值（如公式P(i,j)=Σ_y I(x_i,y,z_j)），从而在粗分辨率点云上形成动态范围更大的图像。室内工业机器人单元和室外数据集（含恶劣天气）的实时多人员检测实验表明，位置精度在体素尺寸（0.04 m）内，室外测试的召回率为76.9%，F1分数为0.87，精度达99.9%。未来工作包括优化分割算法、分类身体部件、扩展类别、测试公共数据集以及将实现移植到更高效的ROS节点以提升检测速率。\n\n### 关键要点\n\n- 系统仅使用深度测量，不依赖RGB图像，确保在RGB信息失效时仍可检测。\n- 数据来自六个安装在天花板上的Kinect V2传感器，经压缩、同步、合并生成单一稠密点云。\n- 点云经体素化（4cm³）后，每个体素带强度值，强度反映该体素内点数量与距离补偿。\n- 通过欧几里得距离聚类，并应用高度、宽度等人体约束筛选候选簇。\n- 候选簇在xz和yz平面展平为灰度图像，利用公式P(i,j)=Σ_y I(x_i,y,z_j)生成像素值。\n- 训练数据包含1105张“人”图像和1956张“非人”图像，随机划分60%训练、30%验证、10%测试，并实施旋转、反射和剪切增强。\n- CNN结构包含三个卷积块（16/32/64滤波器）及批归一化、ReLU、最大池化，输入为224×224×3的彩色图像（灰度复制到三通道）。\n- 室内实验验证了检测器在不同姿势和服装下的准确性，测试准确率为88.2%，位置误差均值在0.035-0.039 m之间，高度测量为1.76±0.06 m（实际1.75 m）。\n- 室外数据集测试中，系统在低角度阳光和暴雨等恶劣条件下仍能正确检测多人，召回率76.9%，F1=0.87，精度99.9%。\n- 当前检测速率约1.2 Hz（受限于Matlab实现和临时文件写入），点云流约5 Hz。\n- 适用场景包括工业机器人单元、海上管道处理设施等，尤其适用于RGB数据不可用的安全关键区域。\n- 局限性：欧几里得分割在人员靠近结构物或携带大物件时可能失败；部分身体被遮挡或暗色服装也可能导致漏检；未来需优化分割、分类部件、扩展类别并移植到原生ROS节点。"
}