{
  "paper_id": "5dac346da843dc027468bf754f0d0767",
  "paper_title": "Gaussian-LIC2: LiDAR-Inertial-Camera Gaussian Splatting SLAM",
  "paper_authors": [
    "Xiaolei Lang",
    "Jiajun Lv",
    "Kai Tang",
    "Laijian Li",
    "Jianxin Huang",
    "Lina Liu",
    "Yong Liu",
    "Xingxing Zuo"
  ],
  "publication_date": "2026",
  "publication_venue": "The International Journal of Robotics Research",
  "abstract": "This paper presents a photo-realistic LiDAR-Inertial-Camera Gaussian Splatting SLAM system that simultaneously addresses visual quality, geometric accuracy, and real-time performance. The proposed method performs robust and accurate pose estimation within a continuous-time trajectory optimization framework, while incrementally reconstructing a large-scale 3D Gaussian map using camera and LiDAR data, all in real time. The resulting map enables high-quality real-time novel view rendering of both RGB images and depth maps. To effectively address under-reconstruction in regions not covered by the LiDAR, we employ a lightweight zero-shot depth model that synergistically combines RGB appearance cues with sparse LiDAR measurements to generate dense depth maps. The depth completion enables reliable Gaussian initialization in LiDAR-blind areas, significantly improving system applicability for sparse LiDAR sensors. To enhance geometric accuracy, we fully use sparse but precise LiDAR depths to supervise Gaussian map optimization and accelerate it with carefully designed CUDA-accelerated strategies. Furthermore, we explore how the incrementally reconstructed Gaussian map can improve the robustness of the odometry. By efficiently incorporating photometric constraints from the Gaussian map into the continuous-time factor graph optimization, we demonstrate improved tracking performance, especially under LiDAR degenerated scenarios. We also showcase downstream applications via extending our elaborate system, including video frame interpolation and fast 3D mesh production. To support rigorous evaluation, we collect a dedicated LiDAR-Inertial-Camera dataset featuring ground-truth poses, depth maps, and extrapolated trajectories for assessing out-of-sequence novel view synthesis in large-scale outdoor scenarios. Extensive experiments on both public and self-collected datasets demonstrate the superiority and versatility of our system across LiDAR sensors with varying sampling densities. Both the dataset and code will be made publicly available on the project page.",
  "keywords": [
    "LiDAR-Inertial-Camera SLAM",
    "Multi-Sensor Fusion",
    "Photo-Realistic Dense Mapping"
  ],
  "paper_institutions": [],
  "output_language": "zh",
  "provider": "doubao",
  "model": "glm-5.3",
  "prompt_template": "simple",
  "extracted_at": "2026-09-17T09:32:13.653673Z",
  "template_tag": "simple",
  "summary": "### 摘要\n\n本文针对LiDAR-惯性-相机高斯泼溅SLAM中视觉质量、几何精度与实时性能难以同时兼顾的核心问题展开研究。现有基于辐射场的SLAM系统（如NeRF类iMAP、NICE-SLAM与3DGS类GS-SLAM、SplaTAM、MonoGS）多面向室内RGB-D场景，在剧烈运动、光照变化与纹理缺失的无界室外环境中表现欠佳；而现有LiDAR-惯性-相机高斯SLAM方法（如Gaussian-LIC、GS-LIVM、GS-LIVO）仍存在LiDAR盲区欠重建、忽视几何重建质量、里程计与高斯建图完全解耦等不足。为此，本文提出Gaussian-LIC2：前端采用基于非均匀累积B样条的连续时间紧密耦合LiDAR-惯性-相机因子图优化（融合LiDAR点面残差、IMU因子、偏置随机游走因子及两种可选相机因子），后端以3D高斯增量构建照片级真实地图；引入轻量级零样本深度补全网络SPNet，协同融合RGB外观线索与稀疏LiDAR测量生成稠密深度图，为LiDAR盲区像素提供可靠的高斯初始化；地图优化采用核心机制——渲染损失 L = L_c + γL_d，其中深度项以渲染不透明度对渲染深度归一化（即 L_d = ||(D̂/Ô − D_s)·M_d||₁），纠正增量插入高斯导致的深度低估，并辅以快速瓦片剔除、逐高斯反向传播、稀疏Adam等CUDA加速策略保障实时性；同时将高斯地图导出的光度约束反馈至连续时间因子图，提升LiDAR退化场景下的跟踪鲁棒性。实验在R3LIVE、FAST-LIVO、MCD、M2DGR四个公共数据集及自采集数据集（含真值位姿、真值深度图与序列外轨迹，覆盖Livox Avia、Mid-360、Ouster OS1-64、Velodyne VLP-32C等多种LiDAR模态与稀疏度）上开展，系统以C++/CUDA实现、基于LibTorch与TensorRT部署，运行于NVIDIA RTX 4090（24GB）平台。评估指标涵盖PSNR、SSIM、LPIPS、Depth-L1、表面重建精度/完整度/Chamfer-L1/F-score、APE RMSE及起终点漂移，结果表明系统在序列内与序列外新视角RGB及深度渲染上全面领先，深度L1误差较前作降低62%，建图总耗时短于数据采集时长，验证了深度补全、深度正则化与CUDA加速等创新的有效性。未来方向包括地图紧凑化、几何精度与视觉质量的进一步平衡、超大场景回环闭合及基础模型前馈高斯预测集成。该系统可服务于机器人实时感知、避障、路径规划与混合现实等实际场景，但受地图内存占用、相机曝光波动与滚动快门效应等约束。\n\n### 关键要点\n\n- **系统定位**：Gaussian-LIC2是首个同时兼顾视觉质量、几何精度与实时性能的LiDAR-惯性-相机高斯泼溅SLAM系统，可在实时鲁棒估计位姿的同时构建照片级真实且几何精确的3D高斯地图。\n- **连续时间轨迹优化**：采用两条非均匀累积B样条分别参数化3D旋转与平移，在0.1秒滑动窗口内联合优化LiDAR点面残差、IMU因子、偏置因子与相机因子，实现异步高频多传感器数据的无插值误差紧密融合。\n- **LiDAR盲区补偿**：选用高效紧凑且泛化能力强的SPNet零样本深度补全模型，将稀疏LiDAR深度与RGB图像融合生成稠密深度图，对30×30无有效LiDAR深度的图像块选取最小深度像素反投影为补充点云，缓解稀疏LiDAR下的欠重建问题。\n- **深度正则化建图**：地图优化损失 L = L_c + γL_d 中，深度项将渲染深度除以渲染不透明度进行归一化，纠正SLAM增量插入高斯导致的像素深度低估，γ设为0.005以平衡视觉与几何质量。\n- **CUDA加速策略**：包括快速瓦片剔除（剔除权重低于1/255的弱影响瓦片）、逐高斯反向传播（32个高斯一桶并行处理避免原子冲突）、稀疏Adam、分离高低阶球谐系数、高效可微SSIM、warp级谓词掩码及固定内存异步传输，保障不牺牲质量下的实时性。\n- **高斯地图反馈里程计**：将高斯地图渲染的RGB图像与原始图像的L1+D-SSIM光度损失优化得到的位姿作为约束引入因子图，在LiDAR退化与纹理缺失同时发生时显著提升跟踪鲁棒性。\n- **数据集构建**：自采集数据集使用Mid-360+RealSense D455与Avia+Hikvision两种手持设备，提供真值位姿（连续时间回环SLAM生成）、真值深度图及序列外轨迹，支持序列内与序列外新视角合成评估。\n- **定位性能**：在退化序列上起终点漂移低至0.03m/0.43°，在MCD大规模室外序列上APE RMSE为0.08-0.09m，优于LVI-SAM、R3LIVE、FAST-LIVO2等传统方法及MonoGS、Co-SLAM、SplaTAM等辐射场SLAM方法。\n- **渲染与重建结果**：在公共与自采集数据集上取得最优的序列内与序列外新视角RGB及深度渲染，深度L1误差相比Gaussian-LIC降低62%；表面重建的Chamfer-L1与F-score亦显著优于对比方法。\n- **实时性验证**：所有序列的建图总耗时均短于rosbag时长（如105秒的Lecture_Hall序列建图仅36秒），位姿估计频率10 Hz，深度补全推理仅需10-20ms每帧。\n- **下游应用**：利用连续时间轨迹任意时刻位姿查询与高斯地图任意视角渲染能力，实现视频帧插值（15Hz提升至30Hz）；增大深度正则化权重γ至2并经TSDF融合与Marching Cubes可快速提取精确3D网格。\n- **局限与展望**：当前未约束地图规模导致内存占用较大，视觉质量与几何精度存在权衡，超大场景会累积位姿漂移，未来计划引入回环闭合、压缩高斯数量并结合基础模型进行前馈高斯预测。"
}