{
  "paper_id": "5ca991a615e565c1278bbaa85c89b6ae",
  "paper_title": "3D People Surveillance on Range Data Sequences of a Rotating Lidar",
  "paper_authors": [
    "Csaba Benedek"
  ],
  "publication_date": "",
  "publication_venue": "Pattern Recognition Letters",
  "abstract": "In this paper, we propose an approach on real-time 3D people surveillance, with probabilistic foreground modeling, multiple person tracking and on-line re-identification. Our principal aim is to demonstrate the capabilities of a special range sensor, called rotating multi-beam (RMB) Lidar, as a future possible surveillance camera. We present methodological contributions in two key issues. First, we introduce a hybrid 2D–3D method for robust foreground-background classification of the recorded RMB-Lidar point clouds, with eliminating spurious effects resulted by quantification error of the discretized view angle, non-linear position corrections of sensor calibration, and background flickering, in particular due to motion of vegetation. Second, we propose a real-time method for moving pedestrian detection and tracking in RMB-Lidar sequences of dense surveillance scenarios, with short- and long-term object assignment. We introduce a novel person re-identification algorithm based on solely the Lidar measurements, utilizing in parallel the range and the intensity channels of the sensor, which provide biometric features. Quantitative evaluation is performed on seven outdoor Lidar sequences containing various multi-target scenarios displaying challenging outdoor conditions with low point density and multiple occlusions.",
  "keywords": [
    "rotating multi-beam Lidar",
    "MRF",
    "motion segmentation",
    "re-identification"
  ],
  "paper_institutions": [
    "Distributed Events Analysis Research Laboratory, Institute for Computer Science and Control, Hungarian Academy of Sciences"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "ark-code-latest",
  "prompt_template": "simple",
  "extracted_at": "2026-08-16T19:46:15.913869Z",
  "template_tag": "simple",
  "summary": "### 摘要\n本文提出了一种基于旋转多光束（RMB）激光雷达的实时三维人群监控方法，包括概率前景建模、多目标跟踪和在线重识别。针对该传感器特有的挑战，首先提出混合2D-3D框架进行鲁棒的前景-背景分割，通过动态马尔可夫随机场（DMRF）在2D范围图像上优化能量函数，并利用3D信息消除量化误差、标定非线性及背景抖动（如植被运动）造成的伪影。其次，设计了一个实时行人检测与跟踪模块，结合短时分配（STA）和长时分配（LTA）处理遮挡和目标的暂时离开与重现，重识别特征完全基于激光雷达的距离和强度通道，提取强度直方图和身高直方图作为生物特征。在七个户外序列上进行了定量评估，涵盖不同季节、光照和拥挤场景，验证了方法在低点密度和多遮挡条件下的有效性，前景分割F-rate达到88.9%-95.1%，处理速度15-16fps，完整流程12-13fps，满足实时要求。实验还探讨了将RMB Lidar作为监控摄像头的可行性，并指出了未来研究方向，如处理更多行人或融合更多传感器特征。\n\n### 关键要点\n- 问题背景：传统光摄像监控受光照和遮挡影响，RMB Lidar提供360°视场、不受光照影响的3D点云，适合户外人群监控，但点云稀疏且存在量化误差和背景运动干扰。\n- 核心方法：前景分割采用混合2D-3D策略，在2D范围图像上用DMRF模型（包含数据项、空间平滑项和时间平滑项）求解，再通过3D点的后投影分类标注点级前景标签，有效减少误检和漏检。\n- 背景建模：每个像素用混合高斯模型（MoG）描述距离分布，通过阈值选择背景分量，计算背景概率，并采用在线K均值更新参数，但单独使用MoG噪声较大。\n- 前景能量：前景能量基于空间邻域中具有相似距离的像素核密度估计，使用sigmoid函数加权，显著降低了背景运动引起的虚假检测。\n- 数据映射：点云投影到圆柱范围图像，水平分辨率$S_W$设为期望列数，垂直分辨率等于传感器数，避免了多目标映射歧义，并通过最近点距离赋值和插值处理空像素。\n- 行人分离：将前景点垂直投影到地面网格，基于点计数阈值提取连通分量，并用体积与距离平方的乘积过滤小噪声，得到行人足部位置候选。\n- 短时跟踪（STA）：使用匈牙利算法匹配测量点与卡尔曼滤波预测位置，阈值限制距离，未匹配轨迹标记为不活跃，允许一定帧数的间歇，期间用预测位置更新。\n- 长时重识别（LTA）：提取强度直方图的Bhattacharyya距离和身高直方图峰值作为特征，仅当两者都相似才认为匹配，用于连接中断轨迹和识别重现行人，新对象需初始化积累特征。\n- 状态机管理：跟踪过程采用有限状态机，结合活跃/不活跃和已识别/初始化状态，超过静默阈值后归档或删除对象，LTA匹配成功则合并轨迹。\n- 实验设置：使用Velodyne HDL-64E，64束激光，帧率15-20Hz，场景为220m²庭院，4-8人，距离1-15m，标注100帧作为前景分割GT，评估F-rate和速度。\n- 结果：DMRF在七个监控序列上F-rate达89.0%-95.1%，优于Bas. MoG（38.4-70.4%）、uniMRF（68.3-89.1%）和3D-MRF（78.7-89.7%），处理速度15-16fps，而3D-MRF仅2-7fps。\n- 跟踪评估：在六人场景中LTA重识别正确率97%，在8人场景（Spring2）正确率76.4%，STA错误极少，完整处理速度12-13fps，验证了系统实时性和有效性。"
}