{
  "paper_id": "3321416f9972581507dc3a36daee5523",
  "paper_title": "Accurate and Real-Time 3D Pedestrian Detection Using an Efficient Attentive Pillar Network",
  "paper_authors": [
    "Duy Tho Le",
    "Hengcan Shi",
    "Hamid Rezatofighi",
    "Jianfei Cai"
  ],
  "publication_date": "",
  "publication_venue": "IEEE Robotics and Automation Letters",
  "abstract": "Efficiently and accurately detecting people from 3D point cloud data is of great importance in many robotic and autonomous driving applications. This fundamental perception task is still very challenging due to (i) significant deformations of human body pose and gesture over time and (ii) point cloud sparsity and scarcity for pedestrian objects. Recent efficient 3D object detection approaches rely on pillar features. However, these pillar features do not carry sufficient expressive representations to deal with all the aforementioned challenges in detecting people. To address this shortcoming, we first introduce a stackable Pillar Aware Attention (PAA) module to enhance pillar feature extraction while suppressing noises in point clouds. By integrating multi-point-channel-pooling, point-wise, channel-wise, and task-aware attention into a simple module, representation capabilities of pillar features are boosted while only requiring little additional computational resources. We also present Mini-BiFPN, a small yet effective feature network that creates bidirectional information flow and multi-level cross-scale feature fusion to better integrate multi-resolution features. Our proposed framework, namely PiFeNet, has been evaluated on three popular large-scale datasets for 3D pedestrian Detection, i.e. KITTI, JRDB, and nuScenes. It achieves state-of-the-art performance on KITTI Bird-eye-view (BEV) as well as JRDB, and competitive performance on nuScenes. Our approach is a real-time detector with 26 frame-per-second (FPS).",
  "keywords": [
    "Deep learning for visual perception",
    "recognition",
    "3D object detection",
    "LiDAR",
    "real-time"
  ],
  "paper_institutions": [],
  "output_language": "zh",
  "provider": "doubao",
  "model": "ark-code-latest",
  "prompt_template": "simple",
  "extracted_at": "2026-08-16T19:46:53.985676Z",
  "template_tag": "simple",
  "summary": "### 摘要\n本文面向机器人与自动驾驶中行人检测的实时性与准确性需求，针对人体姿态多变和点云稀疏导致的柱特征表达不足问题，提出了一种高效的柱感知注意力网络PiFeNet。相关工作背景包括PointPillars、TANet等基于柱的检测器，但它们在行人检测上因柱特征信息丢失与多尺度融合不足而受限。本文提出可堆叠的Pillar Aware Attention (PAA)模块，整合多点通道池化、点注意力、通道注意力和任务感知注意力，增强柱特征表达并抑制噪声；同时设计轻量级Mini-BiFPN特征网络，实现双向信息流和多尺度加权融合。在数据集与实现设置上，分别在KITTI（行人类别，IoU=0.5，BEV和3D检测）、JRDB（AP@0.3/0.5/0.7）和nuScenes（行人AP、mAP、NDS）上评估，采用类似PointPillars的基线，从零训练，使用AdamW和One Cycle策略。主要指标与结果：在KITTI BEV上达到mAP 55.90，3D检测mAP 48.60，优于TANet和Frustum-PointPillars等；JRDB 2019上AP@0.5达42.617，JRDB 2022上AP@0.5达39.838，均为最先进水平；nuScenes行人AP 0.745，NDS 0.606，具竞争力。推理速度26 FPS，验证了实时性。进一步探索方向包括处理密集人群和极端姿态，真实场景约束包括点云稀疏、传感器射程和计算资源限制。核心机制：PAA中的注意力公式A_{c/p}=σ(w1(w0δ(F_mean))+w1(w0δ(F_max)))，以及Mini-BiFPN的加权融合公式F_up^2=conv(δ((w1'·F_in^2+w2'·upsample(F_in^3))/(w1'+w2'+ε)))，通过这些机制增强了特征表达和跨尺度融合。\n\n### 关键要点\n- 提出PiFeNet，一种端到端单阶段3D检测器，结合PAA和Mini-BiFPN，实现实时行人检测（26 FPS）。\n- PAA模块包含多点通道池化（平均池化和最大池化），用于保留柱内所有点的信息。\n- 点注意力和通道注意力通过共享MLP生成注意力向量，公式为A_{c/p}=σ(w1(w0δ(F_mean))+w1(w0δ(F_max)))。\n- 任务感知注意力使用动态激活函数A_T^k(F_c)=max(α_k^1·F_c+β_k^1, α_k^2·F_c+β_k^2)，初始化为ReLU。\n- 堆叠两个PAA模块，第一个输出9个任务感知特征与原始特征结合，第二个进一步处理，提升表达能力。\n- Mini-BiFPN通过可学习权重加权融合多尺度特征，实现双向信息流，公式中采用Swish激活函数。\n- 在KITTI上，PiFeNet在BEV检测中mAP 55.90，3D检测mAP 48.60，优于所有现有方法，尤其超过TANet 2.64 mAP（BEV）。\n- 在JRDB 2019和2022测试集上均排名第一，AP@0.5分别达到42.617和39.838。\n- 在nuScenes上行人AP 0.745，NDS 0.606，相比基线PointPillars++提升0.11 AP。\n- 消融实验表明，单独使用点注意力提升2.3% mAP，通道与任务注意力需联合使用；两个PAA模块平衡精度与速度。\n- 加入Mini-BiFPN后，相比基线mAP提升6.23，验证了多尺度融合的有效性。\n- 运行时间分析：总延迟39.16 ms，其中PAA模块18.12 ms，Mini-BiFPN 13.26 ms，适合实时应用。",
  "available_templates": [
    "deep_read",
    "simple"
  ]
}