文献类型：期刊文献

中文题名：室内环境下融合G-ICP与三维高斯溅射的视觉SLAM算法

英文题名：Visual SLAM Algorithm Integrating G-ICP and 3D Gaussian Splatting for Indoor Environments

作者：张建树[1,2];张军[1,2]

第一作者：张建树

机构：[1]北京联合大学北京市信息服务工程重点实验室,北京100101;[2]北京联合大学机器人学院,北京100101

第一机构：北京联合大学北京市信息服务工程重点实验室

年份：2026

卷号：20

期号：6

起止页码：1627-1636

中文期刊名：计算机科学与探索

外文期刊名：Journal of Frontiers of Computer Science and Technology

收录：;北大核心:【北大核心2023】；

基金：国家自然科学基金(62371013);北京市属高等学校高水平科研创新团队建设支持计划项目(BPHR20220121)。

语种：中文

中文关键词：室内环境;视觉即时定位与建图(SLAM);三维高斯溅射(3DGS);广义迭代最近点(G-ICP)

外文关键词：indoor environments;visual simultaneous localization and mapping(SLAM);3D Gaussian splatting(3DGS);generalized iterative closest point(G-ICP)

摘要：针对传统视觉即时定位与建图(SLAM)算法在室内环境下受高反射物体与低纹理区域的影响,导致定位精度下降与建图质量降低的问题,提出了融合3D高斯溅射(3DGS)与广义迭代最近点(G-ICP)的视觉SLAM算法,命名为GICP-STAM。融合3DGS和G-ICP算法进行位姿初始化;基于G-ICP配准结果进行关键帧筛选,剔除低信息密度的关键帧;提出λ_(SSIM)损失策略进行高斯修剪与致密化,用以滤除异常地图点与创建新地图点。使用三个室内环境的公共数据集进行实验验证。实验结果表明,相较于基线算法SplaTAM,绝对轨迹误差均方根误差(ATE RMSE)在Replica、TUM-RGBD、ScanNet三个数据集上分别提升了38%、11%、3%,在Replica与ScanNet数据集中的平均峰值信噪比(PSNR)分别提升了6%、23%。在室内环境下定位精度与建图质量明显优于基准算法。
To address the problem that traditional visual simultaneous localization and mapping(SLAM)algorithms in indoor environments are affected by highly reflective objects and low-texture regions,resulting in decreased localization accuracy and degraded mapping quality,a visual SLAM algorithm integrating 3D Gaussian splatting(3DGS)and generalized iterative closest point(G-ICP)is proposed,named GICP-STAM.Firstly,3DGS and G-ICP algorithms are integrated for pose initialization.Secondly,keyframe selection is performed based on G-ICP registration results to remove keyframes with low information density.Finally,aλ_(SSIM)loss strategy is proposed for Gaussian pruning and densification to filter out abnormal map points and create new map points.Experimental validation is conducted using three public datasets of indoor environments.The experimental results show that,compared with the baseline algorithm SplaTAM,the absolute trajectory error root mean square error(ATE RMSE)is improved by 38%,11%,and 3%on the Replica,TUM-RGBD,and ScanNet datasets,respectively;the average peak signal-to-noise ratio(PSNR)on the Replica and ScanNet datasets is improved by 6%and 23%,respectively.The localization accuracy and mapping quality in indoor environments are significantly better than those of the baseline algorithm.