← Back ICRA 2026

VS-Graphs: Environment-Aware 3D Scene Graphs for Visual SLAM

Ali Tourani, Saad Ejaz, Miguel Fernandez-Cortizas, Jose Luis Sanchez-Lopez, Holger Voos

PDF

AI summary

Key figure (auto-extracted from paper)

Tightly coupling visual SLAM with online hierarchical 3D scene graph generation improves trajectory accuracy and produces structurally coherent semantic maps in real-time.

Visual SLAM 3D Scene Graphs Semantic Mapping RGB-D Real-time Localization Hierarchical Reasoning

Problem

Most VSLAM systems produce geometric maps that lack semantic interpretation and struggle to model layout-driven structures like walls and rooms, while existing 3D scene graph methods are typically offline, require complete maps, or depend on pre-placed markers.

Approach

vS-Graphs extends ORB-SLAM3 with two online threads that use RGB-D visual and depth cues to detect building components (walls, floors) and infer higher-level structural elements (rooms, floors), integrating them directly into the SLAM optimization loop.

Key results

15.22% average reduction in Absolute Trajectory Error compared to baseline
Lower mapping error with fewer reconstructed points
Semantic detection performance comparable to LiDAR-based systems using only RGB-D input
Real-time incremental construction of an optimizable hierarchical 3D scene graph

Why it matters

Enables robots to navigate and reason about complex indoor environments with richer, structurally meaningful maps without requiring specialized hardware or offline processing.

Abstract

No abstract on file.

Index terms

Semantic Scene Understanding Visual-Inertial SLAM Localization