Rad-GS: Radar-Vision Integration for 3D Gaussian Splatting SLAM in Outdoor Environments

TL;DR

Rad-GS tackles robust outdoor mapping by fusing 4D mmWave radar with monocular vision through a differentiable 3D Gaussian representation, enabling dynamic-free, large-scale SLAM. A Doppler-guided dynamic removal module, coupled with a global octree-based Gaussian map, suppresses moving-object artifacts while maintaining photorealistic rendering and precise localization. The approach demonstrates competitive or superior performance against camera- and LiDAR-based 3D Gaussian SLAM methods, while achieving significant memory reductions suitable for kilometer-scale environments. The results validate the practicality of radar–vision SLAM for robust outdoor scene reconstruction with efficient, scalable memory management.

Abstract

We present Rad-GS, a 4D radar-camera SLAM system designed for kilometer-scale outdoor environments, utilizing 3D Gaussian as a differentiable spatial representation. Rad-GS combines the advantages of raw radar point cloud with Doppler information and geometrically enhanced point cloud to guide dynamic object masking in synchronized images, thereby alleviating rendering artifacts and improving localization accuracy. Additionally, unsynchronized image frames are leveraged to globally refine the 3D Gaussian representation, enhancing texture consistency and novel view synthesis fidelity. Furthermore, the global octree structure coupled with a targeted Gaussian primitive management strategy further suppresses noise and significantly reduces memory consumption in large-scale environments. Extensive experiments and ablation studies demonstrate that Rad-GS achieves performance comparable to traditional 3D Gaussian methods based on camera or LiDAR inputs, highlighting the feasibility of robust outdoor mapping using 4D mmWave radar. Real-world reconstruction at kilometer scale validates the potential of Rad-GS for large-scale scene reconstruction.

Figures (9)

• Figure 1: Overview of Rad-GS: The system comprises a dynamic object removal module, followed by a Gaussian map construction that relies on tracking and map refinement. An octree-based management strategy employs adaptive merging and pruning for Gaussian primitives, yielding a coherent pipeline that transforms raw 4D radar and image data of a kilometer-scale dynamic environment into a memory-efficient, dynamic-free static 3D Gaussian map.
• Figure 2: Doppler-guided dynamic object removal process. (a) Utilize self-motion estimation to detect dynamic points and initialize the octree. (b) Propagate dynamic points and octree nodes to the enhanced radar point cloud. (c) Project octree cells onto the image plane for dynamic object segmentation.
• Figure 3: Effect of Roughness Restriction: Top: Render images. Bottom: Magnified details. The isotropic constraint shows blurred edges of buildings, the normal loss-guided Gaussian primitive rendering sacrifices the rendering of non-planar surfaces, while our loss function achieves the best balance between structural fidelity and texture preservation.
• Figure 4: Illustration of the visual improvement through incremental global optimization. The initialized global 3D Gaussian map (left) and the refined representation with enhanced geometric fidelity and texture realism (right).
• Figure 5: Comparison of dynamic object removal.