{
  "paper_id": "efd7608db5eedf72556ec7d354065621",
  "paper_title": "TimePillars: Temporally-Recurrent 3D LiDAR Object Detection",
  "paper_authors": [
    "Ernesto Lozano Calvo",
    "Bernardo Taveira",
    "Fredrik Kahl",
    "Niklas Gustafsson",
    "Jonathan Larsson",
    "Adam Tonderski"
  ],
  "publication_date": "",
  "publication_venue": "unknown",
  "abstract": "Object detection applied to LiDAR point clouds is a relevant task in robotics, and particularly in autonomous driving. Single frame methods, predominant in the field, exploit information from individual sensor scans. Recent approaches achieve good performance, at relatively low inference time. Nevertheless, given the inherent high sparsity of LiDAR data, these methods struggle in long-range detection (e.g. 200m) which we deem to be critical in achieving safe automation. Aggregating multiple scans not only leads to a denser point cloud representation, but it also brings time-awareness to the system, and provides information about how the environment is changing. Solutions of this kind, however, are often highly problem-specific, demand careful data processing, and tend not to fulfil run-time requirements. In this context we propose TimePillars, a temporally-recurrent object detection pipeline which leverages the pillar representation of LiDAR data across time, respecting hardware integration efficiency constraints, and exploiting the diversity and long-range information of the novel Zenseact Open Dataset (ZOD). Through experimentation, we prove the benefits of having a precurrency, and show how basic building blocks are enough to achieve robust and efficient results.",
  "keywords": [
    "LiDAR",
    "3D object detection",
    "temporal recurrence",
    "PointPillars",
    "Zenseact Open Dataset",
    "long-range detection",
    "convGRU"
  ],
  "paper_institutions": [
    "Chalmers University of Technology",
    "Zenseact"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "ark-code-latest",
  "prompt_template": "simple",
  "extracted_at": "2026-08-16T19:46:12.989700Z",
  "template_tag": "simple",
  "summary": "### 摘要\n本方法针对LiDAR点云稀疏性导致长距离（如200米）目标检测困难的问题，提出TimePillars，一种基于时间递归的3D物体检测流水线。它采用柱状表示（Pillars）并通过卷积GRU（convGRU）记忆单元融合多帧信息，同时在特征层面实现基于卷积的自我运动补偿，并辅以辅助学习任务确保正确性。在Zenseact Open Dataset（ZOD）上，TimePillars与单帧和多帧PointPillars基线比较，显著提升平均精度（AP），尤其在行人和自行车等困难类别及长距离上表现突出，且推理速度远快于多帧聚合方法，满足实时性要求。核心机制是利用convGRU的门控机制和卷积操作，在简单2D卷积基础上实现高效时间融合，并通过实验验证了记忆单元放置于骨干网络之后的最佳配置。\n\n### 关键要点\n- TimePillars采用动态体素化处理点云，避免传统pillar的截断或填充损失，输入张量形状为(N_t, D)，N_t=200000，减少了信息丢失。\n- 网络架构包含pillar特征编码器、2D CNN骨干和检测头；检测头简化了SSD，不用锚框，直接预测每格类别和位置，角度回归采用正弦余弦独立预测。\n- 记忆单元置于骨干网络之后，并采用convGRU，通过合并相似卷积操作提高效率；GRU比LSTM参数更少，起到记忆正则化作用。\n- 自我运动补偿在特征层面用单卷积层实现，输入为相对变换矩阵的6个2D参数（旋转和位移），并通过辅助任务监督训练，辅助任务仅在训练时使用。\n- 在ZOD上，TimePillars（3次扫描）达到NDS 0.728、mAP 0.577，超过单帧PointPillars（NDS 0.677）和多帧（MF）PointPillars（NDS 0.708）。\n- 长距离检测（100-200米以上）中，TimePillars在车辆、自行车和行人类别上均大幅领先单帧模型，如车辆AP从0.498提升到0.591，行人从0.0提升到0.032。\n- 推理速度上，TimePillars（21.5 Hz）比MF PointPillars（9.1 Hz）快约2.3倍，接近单帧（46.2 Hz），但精度显著提高。\n- 消融实验表明，将记忆单元置于骨干网络之后优于之前（NDS 0.727 vs 0.687），且加入辅助任务能提升全部指标（NDS从0.720升至0.727）。\n- 训练采用迁移学习，先训练单帧组件再冻结部分权重训练递归部分；使用随机历史帧数避免过拟合，并利用未标注的过去10帧构建隐藏状态。\n- 方法仅使用2D卷积，兼容Nvidia Orin DLA等目标硬件，避免了稀疏卷积或注意力等不支持的运算。\n- 局限性：仅使用LiDAR，未融合其他传感器；基于单一基线（PointPillars），但框架具有通用性。"
}