{
  "paper_id": "6675b1e3eca99f96a8745e6ed3550edb",
  "paper_title": "Accurate 3-D Semantic Segmentation of Point Clouds for Intelligent Vehicles Based on Multiview Edge Guidance and Fusion",
  "paper_authors": [
    "Yan Liu",
    "Lei Xu",
    "Weiming Hu",
    "Xiong Chen",
    "Bo Yi",
    "Qiu Mao",
    "Dong Kong",
    "Shengping Ruan"
  ],
  "publication_date": "",
  "publication_venue": "IEEE Sensors Journal",
  "abstract": "Semantic segmentation based on LiDAR plays an important role in the environment perception, path planning, and decision control of unmanned ground systems. However, in challenging environments with complex background interleaving, the high proportion of drivable areas and significant target elements pose serious challenges to segmentation tasks. Existing studies attempt to address these challenges but still grapple with issues such as inter-class semantic conflicts and inaccurate edge segmentation. To address these challenges effectively, a reliable point cloud segmentation network driven by multiview fusion and multiedge guidance is specially customized. Specifically, the network's point cloud multiview range image view (RIV) and bird's eye view (BEV) encoding backbones use lightweight pyramid architectures to extract specific details and semantic features of multiscale point clouds. Subsequently, a dedicated multiview edge guidance module focuses on enhancing interclass edge differentiation features from RIV and BEV, respectively. These features are fused with the encoding backbone hierarchy, and the extraction of edge information is supervised by a new hybrid edge loss to maximize the consistency of semantic segmentation and predicted edges. Furthermore, a multiview and multiscale feature fusion module based on a multihead attention mechanism is introduced to enhance the potential complementarity and interaction of the captured features. Through comprehensive experiments and ablation studies on the SemanticKITTI and RELLIS-3D datasets, recognized in the field of autonomous driving, the results demonstrate that our custom method is competitive in terms of intersection over union (IoU) accuracy and real-time indicators compared with typical baseline methods.",
  "keywords": [
    "3-D semantic segmentation",
    "complex environment understanding",
    "multiedge guidance",
    "multiview fusion"
  ],
  "paper_institutions": [],
  "output_language": "zh",
  "provider": "doubao",
  "model": "glm-5.3",
  "prompt_template": "simple",
  "extracted_at": "2026-09-17T09:33:06.152663Z",
  "template_tag": "simple",
  "summary": "### 摘要\n针对无人地面系统在复杂背景交错、可行驶区域占比高且目标要素显著的环境下，LiDAR点云语义分割面临类间语义冲突与边缘分割不准确的问题，本文提出一种由多视图融合与多边缘引导驱动的可靠点云分割网络。现有研究（点基、体素基与投影基方法）虽已取得进展，但仍受制于点云稀疏性、边界模糊及多视图特征交互不足等缺陷。所提网络采用轻量级金字塔架构分别对距离图像视图（RIV）与鸟瞰图（BEV）编码主干提取多尺度点云细节与语义特征，并设计多视图边界引导模块（BGM），其核心机制是基于Hausdorff距离计算子图边界与真值边界的最大差异并取倒数作为相似度（Similarity = 1/max(d(TB,FB), d(FB,TB))），经softmax加权突出类间边缘差异特征；同时引入基于多头注意力机制的多视图多层级交互模块（MMIM）实现BEV与RIV特征级双向深度融合，并以混合边缘损失（Dice边界损失+Lovasz-softmax+交叉熵）监督边缘提取以最大化语义分割与预测边缘的一致性。实验在自动驾驶领域公认的SemanticKITTI（结构化城市场景）与RELLIS-3D（非结构化越野场景）数据集及自采实车数据集上开展，之所以选择这些数据集是为了同时覆盖结构化与非结构化场景以验证方法的适应性与泛化性；实现采用PyTorch，在Ubuntu 20.04系统、单块GeForce RTX 3090 GPU（24GB）上运行，RIV输入尺寸为2048×64、BEV为320×480，SGD优化器（动量0.9、权重衰减0.001）、初始学习率0.01、训练120个epoch。以mIoU和FPS为评价指标，方法在SemanticKITTI上达到68.6% mIoU（较单视图最优LENet提升4.4%、较GFNet提升3.2%），在RELLIS-3D上达到44.8% mIoU（较SalsaNext提升2.3%），消融实验表明BGM与MMIM分别带来约0.6%与0.5%的mIoU增益，验证了多视图边界引导与特征交互融合创新的有效性。未来工作将聚焦于提升小目标分割精度并解决RIV与BEV投影导致的点云信息损失问题。该方法适用于无人系统环境感知、路径规划与决策控制等真实场景，但需在计算资源受限的车载平台上权衡精度与实时性，且投影操作不可避免地造成几何信息损失。\n\n### 关键要点\n- 提出多视图融合与边界引导驱动的3-D点云语义分割网络，据称首次将多视图边界引导准则与跨视图编码结构用于融合多视图点云数据的环境理解任务。\n- 网络包含BEV与RIV双编码分支，均采用轻量级金字塔架构（ResNet-34编码器）提取多尺度点云细节与语义特征，兼顾分割精度与计算成本。\n- 多视图边界引导模块（BGM）将投影图像按通道扩展并分割为N个子图，利用Hausdorff距离度量子图边界与真值边界的最大差异并取其倒数作为相似度，经softmax加权突出边界特征。\n- RIV分支通过球面投影生成距离图像，并引入通道注意力（CA）与高效通道注意力（ECA）协同结构，提升对远距离目标位置信息与跨通道信息交换的感知能力。\n- 多视图多层级交互模块（MMIM）以BEV特征为查询、RIV特征为键值（及反向），通过三个注意力头的多头注意力机制实现双向特征级深度融合，弥补单一投影的信息损失并捕获长程依赖。\n- 混合损失函数由BEV与RIV的Dice边界损失以及Lovasz-softmax与交叉熵分割损失三部分求和构成，监督边缘预测与语义分割的一致性。\n- 实验基于SemanticKITTI（Velodyne HDL-64E采集的结构化道路场景）、RELLIS-3D（Ouster OS1-64采集的13,556帧越野数据，按7800/2413/3343划分训练/验证/测试集，使用14类标注）及Velodyne 32线自采实车数据集，覆盖结构化与非结构化场景。\n- 实现采用PyTorch，运行于Ubuntu 20.04系统与单块RTX 3090 GPU（24GB显存），RIV输入2048×64、BEV输入320×480，SGD优化器动量0.9、权重衰减0.001，初始学习率0.01，训练120个epoch。\n- 在SemanticKITTI上取得68.6% mIoU与11 FPS，较单视图最优方法LENet提升4.4%、较GFNet提升3.2%、较SalsaNext提升约3%，仅比Cylinder3D低0.9%但推理速度更具优势。\n- 在RELLIS-3D上取得44.8% mIoU与14 FPS，较SalsaNext提升2.3%，在草地、树木、灌木等复杂凸场景元素上较Cylinder3D分别提升约1%、3.2%和2.4%。\n- 消融实验表明：相比单RIV输入，双视图输入提升mIoU约2.4%；BGM带来约0.6%提升（额外10.29 GFLOPs）；MMIM带来约0.5%提升（额外12.2 GFLOPs），整体增加24.29 GFLOPs计算量。\n- 方法适用于无人系统环境感知、路径规划与决策控制等真实场景，但投影操作不可避免导致点云几何信息损失，未来需提升小目标分割精度并缓解投影信息损失问题。"
}