{
  "paper_id": "938b163a2fe4aa59e5e07afa06c7c06e",
  "paper_title": "DAPS3D: Domain Adaptive Projective Segmentation of 3D LiDAR Point Clouds",
  "paper_authors": [
    "Alexey A. Klokov",
    "Di Un Pak",
    "Aleksandr Khorin",
    "Dmitry A. Yudin",
    "Leon Kochiev",
    "Vladimir D. Luchinskiy",
    "Vitaly D. Bezuglyj"
  ],
  "publication_date": "2023",
  "publication_venue": "IEEE Access",
  "abstract": "LiDARs are one of the key sources of reliable environmental ranging information for autonomous vehicles. However, segmentation of 3D scene elements (roads, buildings, people, cars, etc.) based on LiDAR point clouds has limitations. On the one hand, point- and voxel-based segmentation neural networks do not offer sufficiently high speed. On the other hand, modern labeled datasets primarily consist of street scenes recorded for driverless cars and contain little data for mobile delivery robots or cleaners that must work in parks and yards with heavy pedestrian traffic. This article aims to overcome these limitations. We have proposed a novel approach called DAPS3D to train deep neural networks for 3D semantic segmentation. This approach is based on a spherical projection of a point cloud and LiDAR-specific masks, enabling the model to adapt to different types of LiDAR. First of all, we have introduced various high-speed multi-scale spherical projection segmentation models, including convolutional, recurrent, and transformer architectures. Among them, the SalsaNextRecLSTM architecture with recurrent blocks showed the best results. In particular, this model achieved the 83.5% mIoU metric for the SemanticKitti dataset with joint categories. Secondly, we have proposed several original augmentations for spherical projections of LiDAR data, including FoV, flip, and rotation augmentation, as well as a special T-Zone cutout. These augmentations increase the model's invariance when dealing with changes in the data domain. Finally, we introduce a new method to generate synthetic datasets for domain adaptation problems. We have developed two new datasets for validating 3D scene outdoor segmentation algorithms: the DAPS-1 dataset, which is based on the augmentation of the reconstructed 3D semantic map, and the DAPS-2 LiDAR dataset, collected by the on-board sensors of a cleaning robot in a park area. Particular attention is given to the performance of the developed models, demonstrating their ability to function in real-time. The code and datasets used in this study are publicly available at: github.com/subake/DAPS3D.",
  "keywords": [
    "Autonomous vehicles",
    "computer vision",
    "artificial neural networks",
    "projection algorithm",
    "image segmentation"
  ],
  "paper_institutions": [
    "Federal Research Center \"Computer Science and Control\" of the Russian Academy of Sciences",
    "Autonomics.tech",
    "Moscow Institute of Physics and Technology",
    "Artificial Intelligence Research Institute (AIRI)"
  ],
  "output_language": "zh",
  "provider": "doubao",
  "model": "glm-5.3",
  "prompt_template": "simple",
  "extracted_at": "2026-09-17T09:32:06.852089Z",
  "template_tag": "simple",
  "summary": "### 摘要\n本文针对自动驾驶与移动机器人中LiDAR点云三维语义分割的两大局限——基于点与体素的神经网络速度不足、现有标注数据集主要面向无人驾驶街道场景而缺乏公园等行人密集环境下配送机器人与清洁机器人的数据——提出了域自适应投影分割方法DAPS3D。与SalsaNet、SalsaNext、FIDNet等虽具实时潜力但对域偏移高度敏感的投影式方法不同，DAPS3D基于点云球面投影与LiDAR专用掩码使模型适应不同类型LiDAR，其核心机制为分割公式 C_P = M(A(π(P)), θ)（π为球面投影、A为增强集合、M为参数θ的神经网络），并以交叉熵损失优化。作者提出含卷积、循环与Transformer架构的高速多尺度球面投影分割模型，其中带LSTM循环块的SalsaNextRecLSTM表现最佳；同时设计了FoV、翻转、旋转、高度、人体插入与T-Zone裁剪等原创增强以提升域不变性，并提出合成数据集生成方法：基于SemanticKITTI经Kimera-Semantic重建三维语义网格、MakeHuman添加人体模型并在Gazebo中按目标LiDAR配置仿真生成半合成数据集DAPS-1（用于预训练并缩减微调所需真实数据规模），以及由莫斯科VDNH公园清洁机器人车载传感器采集的真实数据集DAPS-2（验证真实域迁移效果），并对RELLIS-3D近距人体漏标问题进行标注增强。实验在SemanticKITTI、扩展RELLIS-3D、DAPS-1与DAPS-2上以四类联合类别（车辆、人体、地表、静态障碍物）的IoU/mIoU为指标，采用PyTorch与RTX 2080 Ti、SGD优化器（SalsaNet系40轮、DDRNet系140轮、SegFormer用ImageNet预训练MiT-B1编码器）训练；SalsaNextRecLSTM在SemanticKITTI上取得83.5% mIoU，相比SalsaNext基线在SemanticKITTI、DAPS-1、DAPS-2上分别提升1.4%、4.6%与6.9–11.6% mIoU，增强集合5使基线提升4.3% mIoU（人体类提升8.5%），域自适应使SalsaNetRecLSTM在DAPS-2上提升逾43%，验证了循环块、增强策略与半合成数据三项创新的有效性；所有模型延迟14.4–30.2毫秒满足100毫秒实时要求。未来工作包括扩充DAPS-2、结合时空特征聚合与循环块、适配边缘设备；该方法适用于自主车辆与地面机器人车载感知系统的真实部署，但需应对不同LiDAR配置、机器人本体遮挡及低分辨率下人体类别识别困难等约束。\n\n### 关键要点\n- DAPS3D方法基于点云球面投影与LiDAR专用掩码，核心分割公式为 C_P = M(A(π(P)), θ)，通过增强集合A实现域自适应，仅需单一LiDAR传感器即可完成三维语义分割。\n- 提出SalsaNetRec、SalsaNetRecLSTM、SalsaNextRecLSTM等含循环块的投影分割模型以从点云序列提取更多信息，其中SalsaNextRecLSTM在SemanticKITTI联合类别上取得83.5% mIoU的最佳结果。\n- 设计了FoV、翻转、旋转、高度、人体插入、掩码噪声/丢弃及T-Zone裁剪等原创增强；增强集合5（FoV+翻转+旋转）使SalsaNet基线提升4.3% mIoU、人体类提升8.5%。\n- DAPS-1半合成数据集基于SemanticKITTI构建：先用Kimera-Semantic重建三维语义网格，用MakeHuman生成人体模型插入场景，再在Gazebo中按目标机器人LiDAR配置仿真采集，该流程可复现于任意机器人与LiDAR设置。\n- DAPS-2为2021年夏莫斯科VDNH公园清洁机器人（3个LiDAR配置）实地采集的真实数据集，涵盖不同行人密度的多个场景，用于验证真实域下的分割质量。\n- 对RELLIS-3D进行标注增强：利用在SemanticKITTI与RELLIS-3D上训练80轮后的模型，将原本标注为空类的近距人体点修正为human类。\n- 实验采用PyTorch实现并在NVIDIA RTX 2080 Ti上训练：SalsaNet/SalsaNext系模型用SGD训练40轮（动量0.9、初始学习率0.05、衰减因子0.99），DDRNet系训练140轮，SegFormer使用ImageNet预训练MiT-B1编码器。\n- 评估时将各数据集类别统一映射为车辆、人体、地表、静态障碍物四类并以IoU/mIoU为指标；SalsaNextRecLSTM相比SalsaNext在SemanticKITTI、DAPS-1、DAPS-2上分别提升1.4%、4.6%、6.9–11.6% mIoU。\n- 域自适应效果显著：在DAPS-2上基于DAPS-1训练的模型远优于基于SemanticKITTI训练的模型，SalsaNetRecLSTM质量指标提升逾43%，DDRNetOC提升23%。\n- 所有模型在RTX2080Ti上延迟为14.4–30.2毫秒，满足缓慢移动地面机器人100毫秒的实时要求，且除SalsaNetRecLSTM（108M参数）外参数量均较小。\n- 适配的DDRNet（含Object Context与Deformable Attention模块）与SegFormer模型质量不及SalsaNet/SalsaNext系方法，因低分辨率难以区分类别，在人体类别上尤为明显。\n- 未来工作包括扩充DAPS-2覆盖更多公园区域、在循环块基础上研究时空特征聚合、以及将模型适配至边缘设备。"
}