{
  "paper_id": "44e5a5b72e4f2e678d1eb16f3c043380",
  "paper_title": "Multi Projection Fusion for Real-time Semantic Segmentation of 3D LiDAR Point Clouds",
  "paper_authors": [
    "Yara Ali Alnaggar",
    "Mohamed Afifi",
    "Karim Amer",
    "Mohamed ElHelw"
  ],
  "publication_date": "",
  "publication_venue": "2021 IEEE Winter Conference on Applications of Computer Vision (WACV)",
  "abstract": "Semantic segmentation of 3D point cloud data is essential for enhanced high-level perception in autonomous platforms. Furthermore, given the increasing deployment of LiDAR sensors onboard of cars and drones, a special emphasis is also placed on non-computationally intensive algorithms that operate on mobile GPUs. Previous efficient state-of-the-art methods relied on 2D spherical projection of point clouds as input for 2D fully convolutional neural networks to balance the accuracy-speed trade-off. This paper introduces a novel approach for 3D point cloud semantic segmentation that exploits multiple projections of the point cloud to mitigate the loss of information inherent in single projection methods. Our Multi-Projection Fusion (MPF) framework analyzes spherical and bird's-eye view projections using two separate highly-efficient 2D fully convolutional models then combines the segmentation results of both views. The proposed framework is validated on the SemanticKITTI dataset where it achieved a mIoU of 55.5 which is higher than state-of-the-art projection-based methods RangeNet++ and PolarNet while being 1.6x faster than the former and 3.1x faster than the latter.",
  "keywords": [
    "semantic segmentation",
    "3D point cloud",
    "LiDAR",
    "multi-projection fusion",
    "real-time segmentation",
    "fully convolutional neural networks",
    "autonomous driving"
  ],
  "paper_institutions": [
    "Center for Informatics Science, Nile University"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "glm-5.3",
  "prompt_template": "simple",
  "extracted_at": "2026-09-17T09:25:56.749148Z",
  "template_tag": "simple",
  "summary": "### 摘要\n本文针对自动驾驶汽车与无人机等移动平台上LiDAR点云的实时语义分割问题，提出了多投影融合框架。以往高效的最先进方法（如RangeNet++、PolarNet、SqueezeSeg）依赖单一球面投影将三维点云输入二维全卷积神经网络以平衡精度与速度，但单一投影不可避免地造成信息损失，尤其影响远距离点的分割；而基于三维卷积的方法计算开销过高，不适合实时系统。MPF框架将同一三维点云分别投影为球面视图与鸟瞰视图，由两个独立的高效二维全卷积模型分别处理：球面视图模型以MobileNetV2为编码器骨干并配以转置卷积解码器，鸟瞰视图模型采用轻量化修改版UNet编码器-解码器架构。两分支的分割结果经后处理（以高斯函数 exp(-d²/2σ²) 计算距离权重，对窗口内非稀疏像素的softmax概率进行软投票）映射回原始点云，再将两个视图的得分向量直接相加，取得分最高的类别作为最终预测标签。实验在SemanticKITTI数据集（约4.3万帧360°激光扫描，序列08作为验证集，其余10个序列用于训练）上进行，采用Focal+Lovász-Softmax组合损失、SGD优化器及多种数据增强；结果显示MPF在测试集上取得55.5的mIoU，高于RangeNet++与PolarNet，同时分别比二者快1.6倍和3.1倍，且参数量仅3.18M，验证了多投影融合能有效弥补单一投影的信息损失并保持实时性能。消融实验表明融合使验证集mIoU从50.7提升至57.0（2048分辨率），数据增强带来7.2%的整体提升。未来方向包括将两种投影整合进单一多视图统一模型以及探索更多投影。该框架适用于车载/机载GPU等计算资源受限的实时感知场景，其可扩展性允许根据平台选择或分离部署各分支模型。\n\n### 关键要点\n- 提出多投影融合（MPF）框架，同时利用球面投影与鸟瞰（bird's-eye）投影处理同一LiDAR点云，以缓解单一投影方法固有的信息损失问题，并保持实时性能。\n- 框架包含两个独立分支：球面视图模型基于MobileNetV2骨干加两层转置卷积解码器；鸟瞰视图模型采用轻量化修改版UNet（各含2个下采样与2个上采样块），两分支可按平台需求选择或部署在不同GPU上。\n- 球面投影将点从欧氏空间映射到球坐标并离散化为五通道（x、y、z、r、remission）图像，近距离点优先；鸟瞰投影沿z轴折叠点云，每个网格单元保留z值最大的点，形成四通道图像。\n- 后处理采用高斯加权软投票机制：对每个3D点在投影图像上取K×K滑动窗口，权重按 exp(-d²/2σ²) 随距离衰减，仅统计非稀疏像素，最后按贡献像素数归一化；两视图得分向量直接相加后取最高分类别。\n- 在SemanticKITTI数据集（约43,000帧360°扫描，11个训练序列、11个测试序列，序列08作验证）上训练与评估，使用IoU/mIoU作为评价指标。\n- 训练配置：球面模型用Focal+Lovász-Softmax损失、SGD（动量0.9、权重衰减0.0001、批大小8）、余弦退火热重启（5个周期、初始学习率0.05）；鸟瞰模型用交叉熵+Lovász损失、one-cycle策略训练30个epoch。\n- 数据增强（随机旋转、缩放、平移、翻转、CoarseDropout、裁剪等）使球面模型验证mIoU提升2.7%、鸟瞰模型提升7.7%、整体提升7.2%。\n- 在SemanticKITTI测试集上，MPF（64×2048输入）取得55.5 mIoU，超过RangeNet53++（52.2）与PolarNet（54.3），同时扫描速度分别为二者的1.6倍和3.1倍，参数量仅3.18M（对比RangeNet53的50.4M、PolarNet的13.6M）。\n- 消融实验显示，为单一球面投影模型（SqueezeSeg、SqueezeSegV2、RangeNet53++）加入鸟瞰投影融合均显著提升mIoU（如RangeNet53++ 2048分辨率从50.3升至55.4）；融合对近、远距离点均有改善，鸟瞰视图因笛卡尔坐标下空间分辨率不随距离变化而弥补球面投影的远距离性能退化。\n- 定性结果显示MPF在分割远距离物体及稀有类别（person、bicyclist、other-vehicle）上优于RangeNet53++。\n- 后处理消融表明使用曼哈顿距离且σ=1时效果最佳，更大的滑动窗口未提升分数反而降低整体FPS。\n- 约束与展望：框架需在移动GPU上实时运行（在NVIDIA GTX 1080 Ti上测得20.6 scans/sec@2048分辨率），未来可将两投影整合为单一多视图统一模型并探索使用更多投影。"
}