G3M-SLAM: Anchor-Guided Gaussian Memory for UAV-Oriented Dense SLAM with Representation-Level Submap Fusion
TL;DR
G3M-SLAM is proposed, an anchor-guided Gaussian memory framework for UAV-oriented dense SLAM with representation-level submap fusion with recovery from split-agent degradation and compact representation exchange rather than an improvement over full-sequence single-agent processing.
Abstract
Single-UAV dense visual SLAM is often limited by long trajectory accumulation, incomplete local observations, redundant map growth, and onboard computation constraints. Collaborative mapping can distribute a large mission across several local submaps, but dense 3D Gaussian Splatting (3DGS) maps are expensive to exchange and individual Gaussian primitives are not reliable cross-agent matching units. This paper proposes G3M-SLAM, an anchor-guided Gaussian memory framework for UAV-oriented dense SLAM with representation-level submap fusion. Stable geometric anchors organize local Gaussian primitives and form a compact structural interface for submap exchange, overlap recognition, and correction. A hybrid feature-render tracking strategy combines sparse geometric constraints with Gaussian rendering residuals. A generative completion module predicts candidate Gaussians in weakly observed regions, while multi-view geometric verification and an evidence-aware lifecycle mechanism reject unsupported candidates and control redundant map growth. A dual-graph loop bundle adjustment couples the camera pose graph and the anchor memory graph so that corrections can be propagated to anchor-associated Gaussian structures. Experiments on Replica, ScanNet, TUM RGB-D, and EuRoC MAV evaluate local tracking, dense rendering, runtime, and a split-agent fusion protocol. In the latter protocol, fusion reduces ATE from 0.045 m to 0.031 m, translational RPE from 0.030 m to 0.017 m, and rotational RPE from 1.56° to 0.93°. The serialized anchor packet is 0.66 MB, approximately 101.2× smaller than the complete Gaussian submap. On Jetson AGX Orin, the full system processes EuRoC V101 and V103 at 1.49 FPS and 1.42 FPS, respectively, while ATE increases by only 0.001 m relative to the RTX 3090 Ti workstation results. These gains represent recovery from split-agent degradation and compact representation exchange rather than an improvement over full-sequence single-agent processing. The present study therefore evaluates a representation-level fusion interface and does not reproduce the full conditions of a field-deployed decentralized multi-UAV system.