{
  "paper_id": "dec9dee0920c8c6db7b37a540e965614",
  "paper_title": "LoPF: An Online LiDAR-Only Person-Following Framework",
  "paper_authors": [
    "Xiangyu Chen",
    "Jinhao Liu",
    "Jin Wu",
    "Chaoqun Wang",
    "Rui Song"
  ],
  "publication_date": "",
  "publication_venue": "IEEE Transactions on Instrumentation and Measurement",
  "abstract": "This article presents an online light detection and ranging (LiDAR)-only person-following (LoPF) framework, which can be applied in both indoor and outdoor environments with only sparse LiDAR point cloud data as input. The framework consists of two cascade backbone procedures. The former is pedestrian detection, which is developed for classifying a set of human clusters from noisy and sparse LiDAR point clouds. An originally designed voxel feature is integrated for depicting the 3-D information of each potential cluster. A U-shaped convolutional neural network (CNN) is proposed to achieve human cluster detection. Then the latter procedure, pedestrian tracking, is developed to follow the target. Specifically, to achieve long-term person-following, an unscented Kalman filter (UKF), a position filter, and a target reidentification module are combined. The reidentification module contains a specifically designed support vector machine (SVM) target classifier. The developed framework is verified by extensive real-world experiments in outdoor environments using mobile robots. Compared with the state of the art, our method is better in target identification and tracking, realizing better performance in the person-following.",
  "keywords": [
    "Light detection and ranging (LiDAR)",
    "online support vector machine (SVM)",
    "person-following",
    "voxel feature"
  ],
  "paper_institutions": [
    "Shandong University",
    "The Hong Kong University of Science and Technology"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "ark-code-latest",
  "prompt_template": "simple",
  "extracted_at": "2026-08-16T19:46:22.732549Z",
  "template_tag": "simple",
  "summary": "### 摘要\n本文提出了一种仅基于稀疏LiDAR点云的在线行人跟随框架（LoPF），可应用于室内外环境。针对现有视觉方法受光照影响大、纯LiDAR检测精度不足或实时性差等问题，该框架结合传统方法与深度学习的优势，包含级联的行人检测与行人跟踪两个模块。检测模块通过RANSAC和统计滤波去除地面与噪声，利用欧式聚类和dp-means分割点云，提取包括新设计的体素特征在内的五维特征，并送入U形CNN进行行人分类。跟踪模块采用无迹卡尔曼滤波（UKF）结合全局最近邻数据关联进行多目标跟踪，并设计了基于在线SVM的目标重识别模块，以应对目标丢失后的重新捕获。实验中使用四足机器人和轮式机器人，配备VLP-16 LiDAR，在真实户外和室内场景中验证了框架的有效性。与Koide方法和Yan方法相比，本方法在行人检测精度、跟踪持续性、重识别能力等方面表现更优，如在单人和多场景下取得了更低的ID切换次数、更高的PPV和ACC，以及在长距离跟随中持续跟踪无断裂。核心机制包括体素特征变换公式和基于事件触发的在线SVM更新策略，代码已开源。未来工作将融合相机传感以进一步提升检测和重识别性能。\n\n### 关键要点\n- 提出仅使用稀疏LiDAR点云输入的在线行人跟随框架，适用于室内外环境，无需额外传感器或标记。\n- 设计了一种名为体素特征的新特征，利用三维空间划分保留点云簇的分布信息，对点稀疏性不敏感。\n- 采用U形CNN网络处理体素特征，并结合传统2D特征（点数、最小距离、主副平面直方图）进行行人分类，提高了检测精度。\n- 检测流水线包括RANSAC地面去除、统计离群点滤波、欧式聚类和dp-means分簇，有效降低了噪声干扰。\n- 跟踪模块使用无迹卡尔曼滤波（UKF）与全局最近邻数据关联，并加入位置滤波器以增强目标定位。\n- 设计了基于在线SVM的目标重识别模块，通过事件触发机制在目标丢失时激活，利用最新数据对分类器进行自主学习。\n- 在线SVM使用高斯径向基核，采用30秒固定间隔更新，以匹配行人行走速度和检测范围（直径44.36米）。\n- 训练数据包含来自公开数据集和自采数据的6101个点云簇，其中行人簇4447个、物体簇1654个，用于训练CNN。\n- 通过消融实验验证了五特征组合（含体素特征）达到最高检测准确率92.75%，而仅用前四个特征为85.5%、仅用体素特征为88.25%。\n- 对体素特征尺寸进行选择，最终选用8×8×4，在准确率和时间延迟（14.2ms/s）间取得平衡。\n- 在单人和多人场景的离线数据流及真实机器人平台实验中，本方法在ID切换次数、PPV（0.988）和ACC（0.891）等指标上均优于Koide和Yan的方法。\n- 在户外长时间跟随实验中，四足机器人平台振动下仍能稳定跟踪目标，且重识别模块能防止目标丢失后ID变更。"
}