Rotation Averaging for Map Merging and Trajectory Alignment in Global Structure-from-Motion
(2025) In Master’s Theses in Mathematical Sciences FMAM05 20251Mathematics (Faculty of Engineering)
- Abstract
- Merging of maps, such as 3D point clouds obtained from Structure-from-Motion or visual SLAM, can enhance robustness and expand covered areas in applications like multi-robot systems or augmented reality. When direct data fusion is difficult, an alternative approach is to align pose trajectories. This work investigates the use of rotation averaging as a method for indirect map merging, especially for sequential camera setups where loop closures are typically unavailable. Representing initially aligned data as graphs, with absolute rotations as nodes and relative rotations as edges, we incorporate two-view geometric constraints between sequence pairs using PoseLib, and jointly optimize using weighted edge constraints. Both quantitative... (More)
- Merging of maps, such as 3D point clouds obtained from Structure-from-Motion or visual SLAM, can enhance robustness and expand covered areas in applications like multi-robot systems or augmented reality. When direct data fusion is difficult, an alternative approach is to align pose trajectories. This work investigates the use of rotation averaging as a method for indirect map merging, especially for sequential camera setups where loop closures are typically unavailable. Representing initially aligned data as graphs, with absolute rotations as nodes and relative rotations as edges, we incorporate two-view geometric constraints between sequence pairs using PoseLib, and jointly optimize using weighted edge constraints. Both quantitative evaluations and qualitative visualizations are presented. Results suggest that incorporating rotation averaging between sequences could possibly fine-tune alignment precision, although sensitivity to parameter choices and other issues remain. The study concludes with suggestions for further experiments to advance accuracy and robustness, eventually in practical map merging scenarios including integration with supplementary methods and full pose estimation frameworks involving translation. (Less)
- Popular Abstract (Swedish)
- I många moderna tekniker med exempelvis robotar och förstärkt verklighet behövs tillförlitlig rörelsedata och kartor över omgivningen. Detta projekt fokuserar på att förbättra hur enhetens kamerariktningar, i en sekvens över tid, kan finjusteras efter sammanslagning av två eller fler sådana bildsekvenser, som initiellt roterats till att överlappa utan att deformeras. Genom att använda förhållanden mellan par av kameror, samt ett neuralt nätverk för effektivitet, ska den ursprungliga anpassningen förbättras med hjälp av så kallad rotation averaging (eng), en metod med syfte att "balansera" uppskattningarna av kamerornas orienteringar.
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9207721
- author
- Strömberg, Elina LU
- supervisor
- organization
- course
- FMAM05 20251
- year
- 2025
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- trajectory alignment, rotation averaging, map merging, COLMAP, GLOMAP, NetVLAD, PoseLib, Ceres, LaMAR dataset
- publication/series
- Master’s Theses in Mathematical Sciences
- report number
- LUTFMA-3593-2025
- ISSN
- 1404-6342
- other publication id
- 2025:E51
- language
- English
- id
- 9207721
- date added to LUP
- 2025-09-15 11:10:29
- date last changed
- 2025-09-15 11:10:29
@misc{9207721, abstract = {{Merging of maps, such as 3D point clouds obtained from Structure-from-Motion or visual SLAM, can enhance robustness and expand covered areas in applications like multi-robot systems or augmented reality. When direct data fusion is difficult, an alternative approach is to align pose trajectories. This work investigates the use of rotation averaging as a method for indirect map merging, especially for sequential camera setups where loop closures are typically unavailable. Representing initially aligned data as graphs, with absolute rotations as nodes and relative rotations as edges, we incorporate two-view geometric constraints between sequence pairs using PoseLib, and jointly optimize using weighted edge constraints. Both quantitative evaluations and qualitative visualizations are presented. Results suggest that incorporating rotation averaging between sequences could possibly fine-tune alignment precision, although sensitivity to parameter choices and other issues remain. The study concludes with suggestions for further experiments to advance accuracy and robustness, eventually in practical map merging scenarios including integration with supplementary methods and full pose estimation frameworks involving translation.}}, author = {{Strömberg, Elina}}, issn = {{1404-6342}}, language = {{eng}}, note = {{Student Paper}}, series = {{Master’s Theses in Mathematical Sciences}}, title = {{Rotation Averaging for Map Merging and Trajectory Alignment in Global Structure-from-Motion}}, year = {{2025}}, }