GPU Accelerated Pose Graph Optimization for Stereo Visual SLAM

Yalcin H., ANKARALI M. M.

2026 European Control Conference, ECC 2026, Reykjavik, Iceland, 7 - 10 July 2026, pp.2537-2542, (Full Text)

  • Publication Type: Conference Paper / Full Text
  • City: Reykjavik
  • Country: Iceland
  • Page Numbers: pp.2537-2542
  • Middle East Technical University Affiliated: Yes

Abstract

We explore a GPU-accelerated framework for real-time pose graph optimization in stereo camera-based Simultaneous Localization and Mapping (SLAM). As pose graphs increase in size, due to larger sliding windows and the integration of diverse sensor modalities, real-time optimization becomes challenging for traditional CPU-based solvers. Our approach exploits GPU parallelism to efficiently handle large-scale graphs, achieving optimization of graphs with over 10,000 edges. Each iteration takes only 3 milliseconds on an NVIDIA RTX 2060. Additionally, our work includes a Lie Algebra functions with clear documentation to support fast prototyping of GPU-based pose graph solvers. The full implementation is open-sourced at https://github.com/HaktanM/S-PGO.